CN112202762B - Game defense strategy optimization method and system for sensing edge cloud intelligent jamming attack - Google Patents

Abstract

The invention discloses a game defense strategy optimization method and system for intelligent interference attack in sensing edge cloud. The method comprises the following steps: (1) acquiring a transmission power vector distributed to the computing task of an initial sensor equipment cluster head node set; (2) calculating a power distribution vector of the intelligent interference attacker when the game effect of the intelligent interference attacker is maximized according to a Stark-Boolean model; (3) calculating a transmission power distribution vector of the cluster head node of the sensing equipment when the Nash equilibrium point is reached according to a Stark Boolean model; (4) a decision configuration variable is determined. The system comprises an initialization module, an intelligent interference attacker prediction module, a defense strategy decision module and a configuration module. The invention can effectively defend the intelligent interference attacker with learning ability and provides a defense method for resisting the intelligent interference attack.

Description

Game defense strategy optimization method and system for sensing edge cloud intelligent jamming attack

technical field

The invention belongs to the technical field of the Internet of Things, and more particularly, relates to a game defense strategy optimization method and system for sensing edge cloud intelligent interference attack.

Background technique

The computing tasks of sensing devices can be offloaded to edge service nodes through service access points or base station nodes, which greatly reduces the resource consumption of sensing devices and improves service quality for users. However, in an open environment, the computing task offloading process of sensing devices is vulnerable to intelligent jamming attacks.

The sensing edge cloud system integrates sensing capabilities, control, communication and computing capabilities, and has been widely used in the industrial Internet field. In the sensor edge cloud system, the edge service node on the edge side responds to the request of the sensor device node through an open wireless environment, and receives computing tasks from the sensor device node. Considering the complex wireless communication characteristics on the edge side, in the offloading of computing tasks between sensing devices and edge service nodes, especially the intelligent jamming attack on delay-sensitive computing tasks, the performance of edge computing is reduced or task offloading fails. Therefore, the secure communication between the sensing device and the bedbar device cluster head nodes of the edge service faces great challenges.

The sensor device cluster head node acts as a defender, because it is difficult to capture the channel gain of the intelligent jamming attacker, especially the DDoS (Distributed denial of service attack) attack launched by the intelligent jamming attacker on multiple channels, the more channels are attacked, the defense The higher the computational cost, the higher the computational cost to obtain an optimized defense strategy. S. Bhattacharya et al. formalized a zero-sum chase-evasion game to calculate the optimization strategy and conduct an anti-UAV (Unmanned aerial vehicle) jamming attack (Game-theoretic analysis of an aerialjamming attack on a UAV communication network [C].). L. Xiao et al. considered intelligent UAV jamming attackers, which can specify attack types, such as jamming attacks, eavesdropping attacks, and spoofing attacks, and use reinforcement learning-based power allocation strategies to defend against these attacks (User-Centric View of Unmanned Aerial VehicleTransmission Against Smart Attacks.). Y. Xu et al. considered the incomplete channel state information and used the Bayesian Stackelberg game to study the competitive interaction process between UAV users and jamming attackers (A One-LeaderMulti-Follower Bayesian-Stackelberg Game for Anti-Jamming Transmission in UAVCommunication Networks[J].). In the scheme of Y. Xu et al., the defender uses the Stackelberg anti-jamming attack game to evaluate the impact of the observation error of the intelligent jamming attack on the defense performance, and obtains a Nash equilibrium (A One-Leader Multi-Follower Bayesian-Stackelberg Game for Anti - Jamming Transmission in UAV Communication Networks).

The shortcomings of these schemes are as follows:

(1) The proposed method has limited consideration of incomplete channel state information, which makes the choice of the optimal strategy of the game participants face complexity. When the attack and defense strategies of both game participants change, the proposed method does not provide fast reasoning. function to implement defense strategy selection.

(2) Although the proposed solutions design learning-based defense strategies, they do not consider how to defend against intelligent jamming attackers with learning ability.

(3) Multi-channel attacks on computing task offloading by intelligent jamming attackers, the proposed solutions need to solve complex problems, which greatly reduces the defense performance.

SUMMARY OF THE INVENTION

In view of the above defects or improvement requirements of the prior art, the present invention provides a game defense strategy optimization method and system for sensing edge cloud intelligent interference attack, the purpose of which is to predict the attack strategy of the intelligent interference attacker according to the game model, and to target The predicted attack strategy is used to intelligently optimize the defense strategy, thereby solving the unusable technology that the existing technology cannot target the incomplete channel state information, cannot defend against intelligent jamming attackers with learning ability, or the defense plan is too complicated. question.

In order to achieve the above object, according to one aspect of the present invention, a method for optimizing a game defense strategy under intelligent interference attack in a sensing edge cloud is provided, comprising the following steps:

(1) Obtain the transmission power vector P of the initial sensor device cluster head node set allocated to the computing task: P=(P ₁ , P ₂ , . . . , P _m ), where m is the sensor device cluster head node The number of available channel resources;

(2) According to the Stackelberg model, with the device cluster head node as the leader and the intelligent jamming attacker as the follower, according to the channel gain of the n sensor device cluster head nodes attacked by the intelligent jamming attacker vector, calculating when the game utility of the intelligent jamming attacker is maximized, the power allocation vector J _NN of the intelligent jamming attacker to the n channels it attacks, as the power allocation strategy of the intelligent jamming attacker;

(3) According to the Stackelberg model, on the premise that the intelligent jamming attacker adopts the power allocation strategy of step (2), according to the channel gain vectors of the m channels of the cluster head node of the sensing device, calculate the maximum When the game utility of the sensing device cluster head node reaches the Nash equilibrium point, the transmission power distribution vector P _MM of the sensing device cluster head node to m available channels is used as the power distribution strategy of the sensing device cluster head node ;

(4) According to the power allocation strategy P _MM of the sensor device cluster head node obtained in step (3), the decision configuration variable is determined to perform task offloading.

Preferably, in the method for optimizing the game defense strategy attacked by intelligent jamming in the sensing edge cloud, the calculation method of the game utility of the intelligent jamming attacker in step (2) is as follows:

Among them, n is the total number of channels attacked by the intelligent jamming attacker, P is the transmission power vector of the sensor device cluster head node set allocated to the computing task, and J is the transmission power vector of the intelligent jamming attacker J=(J ₁ , _{J 2 , .} The ith channel of the node is used to unload the computing task, otherwise a _s,i =0; h _s,i is the computing task of the upstream computing task of the s th sensor device cluster head node on the ith channel offload link channel gain, P _i is the transmission power of the sensor device cluster head node of the i-th channel, n _0,i is the noise power of the i-th channel, h _J,i is the channel gain of the intelligent jamming attacker in the i-th channel, and J _i is The transmission power of the intelligent jamming attacker in the i-th channel, and γ is the jamming attack cost per unit jamming power of the intelligent jamming attacker.

Preferably, in the method for optimizing the game defense strategy attacked by intelligent jamming in the sensing edge cloud, in step (2), when the game utility of the intelligent jamming attacker is maximized, the power distribution of the intelligent jamming attacker is calculated. Strategy, use deep neural network to establish intelligent attack model, according to the channel of n sensor equipment cluster head nodes attacked by intelligent jamming attacker, its channel gain vector H _s,i =(h _s,1 ,h _s,2 ,.. .,h _s,n ), predicting the power allocation strategy of the intelligent jamming attacker.

Preferably, in the method for optimizing the game defense strategy attacked by intelligent jamming in the sensing edge cloud, the game utility of maximizing the intelligent jamming attacker in step (2) is denoted as:

Among them, J _max is the maximum transmission power of the intelligent jamming attacker, which is a constant.

The intelligent attack model established by the deep neural network includes an input layer, a normalization layer, a fully connected layer, a data shaping layer, a convolution module, a pooling layer group, and an output layer that are sequentially connected;

The input layer is used to input the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s of the channel of the n sensor device cluster head nodes attacked by the intelligent jamming attacker _{, n} ) to the normalization layer;

并通过全连接层输入到数据整形层；The normalization layer normalizes the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,n ) of the sensor device cluster head node to obtain a normalized The normalized sensor device cluster head node channel gain vector

And input to the data shaping layer through the fully connected layer;

整形后变换成二维矩阵，输入到卷积层。The data shaping layer is used to pass the output of the fully connected layer through the

After shaping, it is transformed into a two-dimensional matrix and input to the convolutional layer.

The convolution module includes two groups of convolution layers, the two groups of convolution kernels are respectively connected by the Relu linear rectification function, and each group of convolution kernels includes n convolution kernels; the normalized sensor device cluster after shaping The channel gain of the head node, after the first convolution operation is performed through the corresponding convolution kernel, is output by the activation function, and the second convolution operation is performed through the corresponding convolution kernel again, and after the activation function, it is output to the pooling layer group;

The pooling layer group includes n parallel pooling layers and fully connected layers;

The output layer is used to output the intelligent jamming attacker transmission power vector (J _N,1 ,J _N,2 ,...,J _N,n ) of the Nash equilibrium point existing in the Stackelberg game, which is the intelligent jamming Attacker's power allocation strategy.

Preferably, the game defense strategy optimization method for being attacked by intelligent interference in the sensing edge cloud, the intelligent attack model established by the deep neural network is obtained by training according to the following method:

使用梯度下降法训练，逐步反相传播调整权值，智能干扰攻击者通过交互获得传感设备簇头节点的传输功率；智能攻击模型的损失函数表示如下：Randomly initialize the multi-channel training weight vector

The gradient descent method is used for training, and the weights are gradually adjusted by anti-phase propagation. The intelligent interference attacker obtains the transmission power of the cluster head node of the sensing device through interaction; the loss function of the intelligent attack model is expressed as follows:

Among them, α _J,i represents the weight coefficient of the loss function, (1-α _J,i )tanh(|J _i -J _max |) is the regularization term, and the power constraint of the intelligent interference attacker participates in the training,

The weight update equation for training the intelligent attack model is:

where _θJ represents the learning rate.

Preferably, in the method for optimizing the game defense strategy attacked by intelligent interference in the sensing edge cloud, the calculation method of the game utility of the cluster head node of the sensing device in step (3) is as follows:

Among them, m is the total number of channel resources available to the sensor device cluster head node, P is the transmission power vector of the sensor device cluster head node set allocated to the computing task, and J is the transmission power vector of the intelligent jamming attacker J=(J ₁ , _{J 2 , .} The ith channel of the node is used to unload the computing task, otherwise a _s,i =0; h _s,i is the computing task of the upstream computing task of the s th sensor device cluster head node on the ith channel offload link channel gain, P _i is the transmission power of the sensor device cluster head node of the i-th channel, n _0,i is the noise power of the i-th channel, h _J,i is the channel gain of the intelligent jamming attacker in the i-th channel, and J _i is The transmission power of the intelligent jamming attacker in the i-th channel, λ is the transmission cost per unit transmission power of the cluster head node of the sensing device.

Preferably, in the game defense strategy optimization method under intelligent interference attack in the sensing edge cloud, when the calculation maximizes the game utility of the sensing device cluster head node to reach a Nash equilibrium point, the sensing device cluster head For the transmission power allocation strategy of nodes, a deep neural network is used to establish a defense-only model, and the power allocation strategy of the sensor equipment cluster head node is obtained.

Preferably, in the method for optimizing the game defense strategy attacked by intelligent interference in the sensing edge cloud, in step (3), the game utility of maximizing the game utility of the cluster head node of the sensing device is denoted as:

Among them, _Pmax is the maximum transmission power of the sensor device cluster head node, which is a constant.

The intelligent defense model established by using a deep neural network includes an input layer, a normalization layer, a fully connected layer, a data shaping layer, a convolution module, a pooling layer group, and an output layer that are sequentially connected;

The input layer is used to input the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,m ) of the m channels of the sensor device cluster head node to normalize chemical layer;

并通过全连接层输入到数据整形层；The normalization layer normalizes the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,m ) of the sensor device cluster head node to obtain a normalized The normalized sensor device cluster head node channel gain vector

And input to the data shaping layer through the fully connected layer;

整形后变换成二维矩阵，输入到卷积层。The data shaping layer is used to pass the output of the fully connected layer through the

After shaping, it is transformed into a two-dimensional matrix and input to the convolutional layer.

The convolution module includes two groups of convolution layers, the two groups of convolution kernels are respectively connected by the Relu linear rectification function, and each group of convolution kernels includes m convolution kernels; the normalized sensor device cluster after shaping The channel gain of the head node, after the first convolution operation is performed through the corresponding convolution kernel, is output by the activation function, and the second convolution operation is performed through the corresponding convolution kernel again, and after the activation function, it is output to the pooling layer group;

The pooling layer group includes m parallel pooling layers and fully connected layers;

The output layer is used to output the output power vector P _MM =(P _M,1 ,P _M,2 ,...,P _M,m ) of the sensor device cluster head node of the Nash equilibrium point existing in the Stackelberg game, That is, the power allocation strategy of the cluster head node of the sensor device.

Preferably, in the method for optimizing the game defense strategy attacked by intelligent interference in the sensing edge cloud, the intelligent defense model in step (3) is trained and obtained according to the following method:

使用梯度下降法训练，逐步反相传播调整权值，传感设备簇头节点通过交互获得智能干扰攻击者的传输功率；智能攻击模型的损失函数表示如下：Randomly initialize the multi-channel training weight vector

The gradient descent method is used for training, and the weights are gradually adjusted by inverse propagation. The cluster head node of the sensing device obtains the transmission power of the intelligent interference attacker through interaction; the loss function of the intelligent attack model is expressed as follows:

通过对损失函数中的功率求偏导，得到：Among them, α _s represents the weight coefficient of the loss function, which balances the influence of constraints on the training process; (1-α _s )tanh(|PP _max |) is the regularization term, and the power constraint of the sensor device cluster head node participates in the training,

By taking the partial derivative of the power in the loss function, we get:

The weight update equation for training the intelligent defense model is:

where _θs is the learning rate.

According to another aspect of the present invention, there is provided a game defense strategy optimization system attacked by intelligent interference in a sensing edge cloud, including an initialization module, an intelligent interference attacker prediction module, a defense strategy decision module, and a configuration module;

The initialization module is used to obtain the initial sensor device cluster head node set and the transmission power vector P: P=(P ₁ , P ₂ , . . . , P _m ) allocated to the computing task and submit it to the intelligent Interference attacker prediction module, where m is the number of available channel resources of the sensor device cluster head node;

The intelligent jamming attacker prediction module is used to follow the Stackelberg model, with the device cluster head node as the leader and the intelligent jamming attacker as the follower, according to the n sensing devices attacked by the intelligent jamming attacker The channel gain vector of the channel of the cluster head node, when the game utility of the intelligent jamming attacker is maximized, the power allocation vector J _NN of the intelligent jamming attacker to the n channels it attacks is calculated as the power of the intelligent jamming attacker assigning a strategy and submitting it to the defense strategy decision-making module;

The defense strategy decision-making module is used to calculate the maximum value according to the channel gain vectors of the m channels of the cluster head node of the sensing device according to the Stackelberg model and on the premise that the intelligent jamming attacker adopts a power allocation strategy. When the game utility of the sensing device cluster head node is optimized to reach the Nash equilibrium point, the transmission power distribution vector of the sensing device cluster head node to m available channels is used as the power distribution strategy of the sensing device cluster head node, Submit to the configuration module;

The configuration module is configured to determine decision configuration variables according to the power distribution strategy P _MM of the sensor device cluster head node, and perform task offloading. In general, compared with the prior art, the following beneficial effects can be achieved by the above technical solutions conceived by the present invention:

In the sensing edge cloud environment, the present invention aims at the problem of offloading security tasks between sensing equipment and edge service nodes. Based on the Stackelberg game, the sensing equipment cluster head node with sufficient resources optimizes the power allocation strategy and defends against intelligent interference attacks through learning. , which can effectively defend against intelligent jamming attackers with learning ability, and provides a defense method against intelligent jamming attacks.

The preferred solution is based on the learning process of the Stackelberg game strategy based on DNN (Deep neural network), in which the cluster head node of the sensing device plays the role of leader, and the intelligent interference attacker plays the role of follower. First, the intelligent jamming attacker obtains the transmission power allocation strategy of the cluster head node of the sensing device, and learns the optimal power allocation strategy through its own channel gain, and maximizes its game utility. Secondly, the cluster head node of the sensing device obtains the power allocation strategy of the intelligent jamming attacker, and learns the optimal power allocation strategy through its own channel gain to maximize its game utility. make defensive decisions effectively

Description of drawings

1 is a schematic structural diagram of the sensing edge cloud targeted by the present invention;

2 is a schematic structural diagram of an intelligent attack model provided by an embodiment of the present invention;

3 is a schematic structural diagram of a single-channel intelligent attack model provided by an embodiment of the present invention;

4 is a schematic structural diagram of an intelligent defense model provided by an embodiment of the present invention;

FIG. 5 is a schematic structural diagram of a single-channel intelligent defense model provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

In the present invention, the sensing edge cloud system is composed of sensing equipment nodes, sensing equipment cluster head nodes, edge service nodes and intelligent jamming attackers. Among them, the edge service node consists of AP access points and micro cloud services. As shown in Figure 1. When attacked by intelligent jamming, the sensing device node offloads its computing tasks to the edge service node through the sensing device cluster head node, and uses the computing power of the sensing device cluster head node to implement defense strategies against intelligent jamming attackers with learning ability.

The game defense strategy optimization method for being attacked by intelligent interference in the sensing edge cloud provided by the present invention includes the following steps:

(1) Obtain the transmission power vector P of the initial sensor device cluster head node set allocated to the computing task: P=(P ₁ , P ₂ , . . . , P _m ), where m is the sensor device cluster head node The number of available channel resources;

(2) According to the Stackelberg model, with the device cluster head node as the leader and the intelligent jamming attacker as the follower, according to the channel gain of the n sensor device cluster head nodes attacked by the intelligent jamming attacker vector, calculating the power allocation vector J _NN of the intelligent jamming attacker to the n channels attacked when the game utility of the intelligent jamming attacker is maximized, as the power allocation strategy of the intelligent jamming attacker.

The calculation method of the game utility of the intelligent interference attacker is as follows:

Among them, n is the total number of channels attacked by the intelligent jamming attacker, P is the transmission power vector of the sensor device cluster head node set allocated to the computing task, and J is the transmission power vector of the intelligent jamming attacker J=(J ₁ , _{J 2 , .} The ith channel of the node is used to unload the computing task, otherwise a _s,i =0; h _s,i is the computing task of the upstream computing task of the s th sensor device cluster head node on the ith channel offload link channel gain, P _i is the transmission power of the sensor device cluster head node of the i-th channel, n _0,i is the noise power of the i-th channel, h _J,i is the channel gain of the intelligent jamming attacker in the i-th channel, and J _i is The transmission power of the intelligent jamming attacker in the i-th channel, and γ is the jamming attack cost per unit jamming power of the intelligent jamming attacker.

When the calculation is to maximize the game utility of the intelligent jamming attacker, the power allocation strategy of the intelligent jamming attacker is preferably a deep neural network to establish an intelligent attack model, according to the n sensing device cluster heads attacked by the intelligent jamming attacker. The channel of the node has its channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,n ), which predicts the power allocation strategy of the intelligent jamming attacker; the details are as follows:

The game utility of the maximizing intelligence jamming attacker is written as:

Among them, J _max is the maximum transmission power of the intelligent jamming attacker, which is a constant.

The intelligent attack model established by the deep neural network, as shown in Figure 2, includes an input layer, a normalization layer, a fully connected layer, a data shaping layer, a convolution module, a pooling layer group, and an output layer that are sequentially connected. Floor;

The input layer is used to input the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s of the channel of the n sensor device cluster head nodes attacked by the intelligent jamming attacker _{, n} ) to the normalization layer;

并通过全连接层输入到数据整形层；计算方法如下：The normalization layer normalizes the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,n ) of the sensor device cluster head node to obtain a normalized The normalized sensor device cluster head node channel gain vector

And input to the data shaping layer through the fully connected layer; the calculation method is as follows:

where E(·) is the expectation and D(·) is the variance.

整形后变换成二维矩阵，输入到卷积层。The data shaping layer is used to pass the output of the fully connected layer through the

After shaping, it is transformed into a two-dimensional matrix and input to the convolutional layer.

The convolution module includes two groups of convolution layers, the two groups of convolution kernels are respectively connected by the Relu linear rectification function, and each group of convolution kernels includes n convolution kernels; the normalized sensor device cluster after shaping The channel gain of the head node, after the first convolution operation is performed through the corresponding convolution kernel, is output by the activation function, and the second convolution operation is performed through the corresponding convolution kernel again, and after the activation function, it is output to the pooling layer group;

The first convolution operation uses a 3×3 convolution kernel with a stride of 1 to extract n channel gain information of the intelligent jammer attacker. The output of the convolution operation is as follows:

为权值，第一次卷积运算的输出向量为

in,

is the weight, the output vector of the first convolution operation is

其中

After the first convolution operation, the Relu function is used as the activation function, and the output vector is:

in

第二次卷积运算的输出向量为

The second convolution operation uses a 3×3 convolution kernel with a stride of 1. The output of the convolution operation is as follows:

The output vector of the second convolution operation is

其中

After the second convolution operation, the Relu function is used as the activation function, and the output vector is:

in

The pooling layer group includes n parallel pooling layers and fully-connected layers; preferably, in order to speed up the training speed, the maximum pooling layer is used, and the maximum pooling layer is obtained through a 2×2 sliding window with a step size of 1. The output vector of the layer is

The output layer preferably adopts the Sigmoid function to output the intelligent interference attacker transmission power vector (J _N,1 ,J _N,2 ,...,J _N,n ) of the Nash equilibrium point existing in the Stackelberg game, which is the The power allocation strategy of the intelligent jamming attacker is described; the Sigmoid function is preferably used as the output layer,

The intelligent attack model established by the deep neural network is obtained by training according to the following method:

使用梯度下降法训练，逐步反相传播调整权值，智能干扰攻击者通过交互获得传感设备簇头节点的传输功率；智能攻击模型的损失函数表示如下：Randomly initialize the multi-channel training weight vector

The gradient descent method is used for training, and the weights are gradually adjusted by anti-phase propagation. The intelligent interference attacker obtains the transmission power of the cluster head node of the sensing device through interaction; the loss function of the intelligent attack model is expressed as follows:

通过对损失函数中的功率求偏导，得到：Among them, α _J,i represents the weight coefficient of the loss function, (1-α _J,i )tanh(|J _i -J _max |) is the regularization term, and the power constraint of the intelligent interference attacker participates in the training,

By taking the partial derivative of the power in the loss function, we get:

The weight update equation for training the intelligent attack model is:

where _θJ represents the learning rate.

After the training of the intelligent attack model is completed, the trained intelligent attack model is used to predict the transmission power strategy of the intelligent jammer. The power allocation strategy J _NN = (J _N,1 ,J _N,2 ,...,J _N,n ) of the intelligent jamming attacker output by the intelligent attack model.

(3) According to the Stackelberg model, on the premise that the intelligent jamming attacker adopts the power allocation strategy of step (2), according to the channel gain vectors of the m channels of the cluster head node of the sensing device, calculate the maximum When the game utility of the sensing device cluster head node reaches the Nash equilibrium point, the transmission power distribution vector of the sensing device cluster head node to m available channels is used as the power distribution strategy of the sensing device cluster head node;

The calculation method of the game utility of the cluster head node of the sensing device is as follows:

Among them, m is the total number of channel resources available to the sensor device cluster head node, P is the transmission power vector of the sensor device cluster head node set allocated to the computing task, and J is the transmission power vector of the intelligent jamming attacker J=(J ₁ , _{J 2 , .} The ith channel of the node is used to unload the computing task, otherwise a _s,i =0; h _s,i is the computing task of the upstream computing task of the s th sensor device cluster head node on the ith channel offload link channel gain, P _i is the transmission power of the sensor device cluster head node of the i-th channel, n _0,i is the noise power of the i-th channel, h _J,i is the channel gain of the intelligent jamming attacker in the i-th channel, and J _i is The transmission power of the intelligent jamming attacker in the i-th channel, λ is the transmission cost per unit transmission power of the cluster head node of the sensing device.

When the calculation maximizes the game utility of the sensor device cluster head node to reach the Nash equilibrium point, the transmission power distribution strategy of the sensor device cluster head node is preferably a deep neural network to establish a defense-only model to obtain the sensor device cluster. The power allocation strategy of the head node; the details are as follows:

The game utility of maximizing the cluster head node of the sensing device is denoted as:

Among them, _Pmax is the maximum transmission power of the sensor device cluster head node, which is a constant.

The intelligent defense model established by the deep neural network, the structure is shown in Figure 4, including the sequentially connected input layer, normalization layer, fully connected layer, data shaping layer, convolution module, pooling layer group, and output Floor;

The input layer is used to input the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,m ) of the m channels of the sensor device cluster head node to normalize chemical layer;

并通过全连接层输入到数据整形层；计算方法如下：The normalization layer normalizes the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,m ) of the sensor device cluster head node to obtain a normalized The normalized sensor device cluster head node channel gain vector

And input to the data shaping layer through the fully connected layer; the calculation method is as follows:

where E(·) is the expectation and D(·) is the variance.

整形后变换成二维矩阵，输入到卷积层。The data shaping layer is used to pass the output of the fully connected layer through the

After shaping, it is transformed into a two-dimensional matrix and input to the convolutional layer.

The convolution module includes two groups of convolution layers, the two groups of convolution kernels are respectively connected by the Relu linear rectification function, and each group of convolution kernels includes m convolution kernels; the normalized sensor device cluster after shaping The channel gain of the head node, after the first convolution operation is performed through the corresponding convolution kernel, is output by the activation function, and the second convolution operation is performed through the corresponding convolution kernel again, and after the activation function, it is output to the pooling layer group;

The first convolution operation uses a 3×3 convolution kernel with a stride of 1 to extract m channel gain information of the intelligent interference attacker. The output of the convolution operation is as follows:

为权值，第一次卷积运算的输出向量为

in,

is the weight, the output vector of the first convolution operation is

其中

After the first convolution operation, the Relu function is used as the activation function, and the output vector is:

in

第二次卷积运算的输出向量为

The second convolution operation uses a 3×3 convolution kernel with a stride of 1. The output of the convolution operation is as follows:

The output vector of the second convolution operation is

其中

After the second convolution operation, the Relu function is used as the activation function, and the output vector is:

in

The pooling layer group includes m parallel pooling layers and fully-connected layers; preferably, in order to speed up the training speed, the maximum pooling layer is used, and the maximum pooling layer is obtained through a 2×2 sliding window with a step size of 1. The output vector of the layer is

The output layer is used to output the output power vector P _MM =(P _M,1 ,P _M,2 ,...,P _M,m ) of the sensor device cluster head node of the Nash equilibrium point existing in the Stackelberg game, is the power allocation strategy of the cluster head node of the sensor device; preferably the Sigmoid function is used as the output layer,

Train as follows:

使用梯度下降法训练，逐步反相传播调整权值，传感设备簇头节点通过交互获得智能干扰攻击者的传输功率；智能攻击模型的损失函数表示如下：Randomly initialize the multi-channel training weight vector

The gradient descent method is used for training, and the weights are gradually adjusted by inverse propagation. The cluster head node of the sensing device obtains the transmission power of the intelligent interference attacker through interaction; the loss function of the intelligent attack model is expressed as follows:

通过对损失函数中的功率求偏导，得到：Among them, α _s represents the weight coefficient of the loss function, which balances the influence of constraints on the training process; (1-α _s )tanh(|PP _max |) is the regularization term, and the power constraint of the sensor device cluster head node participates in the training,

By taking the partial derivative of the power in the loss function, we get:

The weight update equation for training the intelligent defense model is:

where _θs is the learning rate.

After the training of the intelligent defense model is completed, the trained intelligent defense model is used to decide the transmission power strategy of the cluster head node of the sensor equipment. When the channel gain vectors of m sensor equipment nodes are input, the sensing The power allocation policy P _MM of the device cluster head node.

(4) According to the power allocation strategy P _MM of the sensor device cluster head node obtained in step (3), the decision configuration variable is determined to perform task offloading.

The game defense strategy optimization system attacked by intelligent interference in the sensing edge cloud provided by the present invention includes an initialization module, an intelligent interference attacker prediction module, a defense strategy decision module, and a configuration module;

The initialization module is used to obtain the initial sensor device cluster head node set and the transmission power vector P: P=(P ₁ , P ₂ , . . . , P _m ) allocated to the computing task and submit it to the intelligent Interference attacker prediction module, where m is the number of available channel resources of the sensor device cluster head node;

The intelligent jamming attacker prediction module is used to follow the Stackelberg model, with the device cluster head node as the leader and the intelligent jamming attacker as the follower, according to the n sensing devices attacked by the intelligent jamming attacker The channel gain vector of the channel of the cluster head node, when the game utility of the intelligent jamming attacker is maximized, the power allocation vector J _NN of the intelligent jamming attacker to the n channels it attacks is calculated as the power of the intelligent jamming attacker assigning a strategy and submitting it to the defense strategy decision-making module;

The defense strategy decision-making module is used to calculate the maximum value according to the channel gain vectors of the m channels of the cluster head node of the sensing device according to the Stackelberg model and on the premise that the intelligent jamming attacker adopts a power allocation strategy. When the game utility of the sensing device cluster head node is optimized to reach the Nash equilibrium point, the transmission power distribution vector of the sensing device cluster head node to m available channels is used as the power distribution strategy of the sensing device cluster head node, Submit to the configuration module;

The configuration module is configured to determine decision configuration variables according to the power distribution strategy P _MM of the sensor device cluster head node, and perform task offloading.

The following are examples:

When the sensor device cluster head node unloads delay-sensitive computing tasks, the intelligent jamming attacker increases the delay time and energy consumption of computing task offloading, and reduces the reliability of the channel and the offloading capacity of computing tasks. Let the decision configuration variable for computing task offloading be A={a _{s, i} |s ∈ M, i ∈ E}, where M represents the cluster head node set of the sensing device, and E represents the channel set of the sensing device cluster head node. If the sensor device cluster head node s utilizes channel i to offload computing tasks, then a _s,i =1, otherwise a _s,i =0. Therefore, for a single channel i, the computing task offload capacity of the sensor device cluster head node s is:

As a defender of anti-jamming attacks, the sensing device cluster head node s selects a certain transmission power to offload computing tasks. Therefore, the game utility of the sensing device cluster head node s is:

Among them, P represents the transmission power of the sensor device cluster head node s, and J represents the transmission power of the intelligent jamming attacker. The power per unit of the sensor device cluster head node and the intelligent jamming attacker are λ and γ, respectively. n ₀ represents the noise power. h _s,i represents the channel gain of the uplink computing task offloading link, and h _J,i represents the channel gain of the intelligent jammer attacker.

The intelligent jamming attacker selects a certain power to jam the offloading process of the delay-sensitive computing task of the cluster head node s of the sensing device. Therefore, the game utility of the intelligent jamming attacker is:

且

m个信道模式下，传感设备簇头节点的博弈效用为：When the cluster head node of the sensing device has m available channel resources, the intelligent jamming attacker initiates a multi-channel attack, which leads to the failure of transmission of offloaded computing tasks on multiple channels. Let P _i and J _i denote the transmission power allocated to channel i by the sensor device cluster head node and the intelligent jamming attacker. Let P=(P ₁ , P ₂ , . . . , P _m ) denote the transmission power vector of the cluster head node of the sensing device. J=(J ₁ , J ₂ , . . . , J _n ) represents the transmission power vector of the intelligent jamming attacker. The transmission power of the sensor device cluster head node and the intelligent jamming attacker satisfies

and

In m channel mode, the game utility of the sensor device cluster head node is:

In the multi-channel mode, the intelligent jamming attacker performs jamming attacks on n channels, and the game utility of the intelligent jamming attacker is:

When attacked by intelligent jamming attackers with learning ability, the present invention models the power allocation problem of anti-jamming attacks as a DNN-based Stackelberg game. In the game model, the sensor device cluster head node and the intelligent jamming attacker are the game participants. The cluster head node of the sensing device is the leader, and the transmission power is allocated first. The intelligent jamming attacker is the follower, which performs jamming attacks on the computing task offloading process of the cluster head node of the sensing device. The optimal power allocation strategy for each game participant is obtained through DNN inference. The invention designs a defense strategy for an intelligent jamming attacker with learning ability. The intelligent jamming attacker can obtain the transmission power of the cluster head node of the sensing device through game interaction, and at the same time infers the power distribution strategy according to its own channel gain to maximize its game utility. . Similarly, as a defender, the cluster head node of the sensing device can also obtain the power allocation strategy of the intelligent interference attacker through game interaction, and at the same time calculate the power allocation strategy of task offloading according to its own channel gain inference.

(1) Obtain the transmission power vector P of the initial sensor device cluster head node set allocated to the computing task: P=(P ₁ , P ₂ , . . . , P _m ), where m is the value of the sensor device cluster node The number of available channel resources;

(2) According to the Stackelberg model, with the device cluster head node as the leader and the intelligent jamming attacker as the follower, according to the channel gain of the n sensor device cluster head nodes attacked by the intelligent jamming attacker vector, calculating when the game utility of the intelligent jamming attacker is maximized, the power allocation vector J _NN of the intelligent jamming attacker to the n channels it attacks, as the power allocation strategy of the intelligent jamming attacker;

When the intelligent jamming attacker launches a multi-channel attack on the offload link, the game utility of maximizing the intelligent jamming attacker can be formalized as follows:

In this embodiment, a multi-channel intelligent attack model network MJnet is established to maximize the multi-channel attack game utility of the intelligent jamming attacker. At the same time, MJnet is trained to infer the optimal attack strategy vector of intelligent jamming attackers in multi-channel mode.

MJnet structure, as shown in Figure 2:

The input and output steps of MJnet processing the channel gain of the intelligent jammer are as follows:

①Input: The channel gain of the normalized intelligent jamming attacker's multi-channel attack is a standard normal distribution with a mean of 0 and a variance of 1. The input to MJnet multi-channel gain normalization is:

整形后变换成n个二维矩阵存储信道增益。② After

After shaping, it is transformed into n two-dimensional matrices to store the channel gains.

(3) The shaped data is input to the convolution layer, and n convolution kernels with a stride of 1 are used to perform the first convolution operation to extract n channel gain information of the intelligent interference attacker.

其中

为权值。因此，第一次卷积的输出向量为

④ The output of the first convolutional layer is:

in

is the weight. Therefore, the output vector of the first convolution is

为relu的单个输出，多输出向量为

⑤ The output of the first convolutional layer uses relu as the activation function,

is the single output of relu, and the multi-output vector is

多输出向量为

⑥ Then perform the second convolution operation, and the single output is:

The multi-output vector is

为relu的单个输出，多输出向量为

⑦ Use relu as the activation function again,

is the single output of relu, and the multi-output vector is

⑧ In order to speed up the training speed, the maximum pooling layer is used, and the output vector of the maximum pooling layer is obtained after a 2×2 sliding window with a step size of 1.

多输出向量为(J_N,1,J_N,2,...,J_N,N)。⑨ Finally, the Sigmoid function is used, and the single output is

The multi-output vector is (J _N,1 ,J _N,2 ,...,J _N,N ).

The MJnet training and inference process is as follows:

初始化完成后，使用随机梯度下降法训练MJnet，逐步反向传播调整权值，使得智能干扰攻击者的博弈效用值最大。其中，智能干扰攻击者通过交互获得传感设备簇头节点的传输功率。智能干扰攻击者进行多信道攻击时最大化其博弈效用的损失函数表示如下：Intelligent jamming attackers reason their multi-channel attack strategies through MJnet. Therefore, the intelligent jamming attacker first randomly initializes the multi-channel training weight vector

After the initialization is completed, the stochastic gradient descent method is used to train MJnet, and the weights are adjusted by back-propagation step by step, so that the game utility value of the intelligent interference attacker is maximized. Among them, the intelligent jamming attacker obtains the transmission power of the cluster head node of the sensing device through interaction. The loss function that maximizes the game utility of an intelligent jamming attacker when conducting a multi-channel attack is expressed as follows:

通过对损失函数中的功率J_i求偏导，得到：Among them, α _J,i represents the weight coefficient of the loss function, (1-α _J,i )tanh(|J _i -J _max |) is the regularization term, and the power constraint of the intelligent interference attacker participates in the training,

By taking the partial derivative of the power J _i in the loss function, we get:

The weight update equation for training MJnet is:

MJnet输出优化的多信道攻击功率策略向量。where _θJ represents the learning rate. After the MJnet training is completed, use the trained MJnet to reason about the transmission power strategy of the multi-channel attack of the intelligent jammer. The multi-channel gain vector of the input smart jammer attacker

MJnet outputs the optimized multi-channel attack power policy vector.

When m=n=1, that is, in the single-channel attack mode, the game utility of maximizing the intelligent jamming attacker can be specially simplified as:

MJnet can be specifically simplified as a single-layer SJnet. By training SJnet to learn and reason the response strategy of the intelligent jamming attacker in the single-channel attack mode, so as to maximize the game utility of the intelligent jamming attacker.

The structure of SJnet is shown in Figure 3: The network model consists of normalization, full connection layer, data shaping, convolution layer, pooling layer, etc.

The input and output steps of SJnet processing the channel gain of the intelligent jammer are as follows:

①In order to speed up the convergence speed of the game strategy learning of the intelligent jamming attacker and ensure that the input channel gain of the intelligent jamming attacker is identically distributed. Therefore, the input channel gain is normalized to a standard normal distribution with mean 0 and variance 1. The normalized output of the smart jammer's channel gain is:

输入SJnet全连接层。where E(·) is the expectation and D(·) is the variance. The normalized channel gain vector of the smart jammer

Enter the SJnet fully connected layer.

整形后变换成二维矩阵。② data shaping. The output of the fully connected layer goes through

After shaping, it is transformed into a two-dimensional matrix.

(3) The shaped data is input to the convolution layer, and the first convolution operation is performed using a 3×3 convolution kernel with a stride of 1, and the channel gain information of the key changes of the intelligent jammer is extracted.

其中

为权值。④ The output of convolutional layer 1 is:

in

is the weight.

为relu的输出。⑤ The output of convolutional layer one uses relu as the activation function,

is the output of relu.

⑥ Then perform the second convolution operation, and the output of the second convolution layer is:

为relu的输出。⑦ Use relu as the activation function again,

is the output of relu.

⑧ In order to speed up the training, use the maximization pooling layer and use a 2×2 sliding window with a stride of 1, and the output of the maximization pooling layer is

⑨ Using a fully connected layer, stretch the output of the maximization pooling layer into an n×1 vector.

⑩ Finally, use the Sigmoid function to output

The SJnet training and inference process is as follows:

初始化完成后，使用随机梯度下降法训练SJnet，逐步反向传播调整权值，使得智能干扰攻击者的博弈效用值最大。其中，智能干扰攻击者通过交互获得传感设备簇头节点的传输功率。智能干扰攻击者最大化其博弈效用的损失函数表示如下：Intelligent jamming attackers reason their attack strategies through SJnet. Therefore, the intelligent jamming attacker first randomly initializes the training weights

After the initialization is completed, the SJnet is trained using the stochastic gradient descent method, and the weights are gradually adjusted by backpropagation, so that the game utility value of the intelligent interference attacker is maximized. Among them, the intelligent jamming attacker obtains the transmission power of the cluster head node of the sensing device through interaction. The loss function for an intelligent jamming attacker to maximize its game utility is expressed as follows:

Among them, α _J represents the weight coefficient of the loss function. The second term is the regularization term, the interference power constraint of the intelligent jamming attacker participates in the training,

By taking the partial derivative of the power J in the loss function, we get:

The weight update equation for training SJnet is:

SJnet输出优化的单信道攻击功率策略。where _θJ represents the learning rate. After the SJnet training is completed, use the trained SJnet to infer the transmission power strategy of the intelligent jammer attacker. When inputting the channel gain vector of the intelligent jammer attacker

SJnet outputs an optimized single-channel attack power policy.

(3) According to the Stackelberg model, on the premise that the intelligent jamming attacker adopts the power allocation strategy of step (2), according to the channel gain vectors of the m channels of the cluster head node of the sensing device, calculate the maximum When the game utility of the sensing device cluster head node reaches the Nash equilibrium point, the transmission power distribution vector of the sensing device cluster head node to m available channels is used as the power distribution strategy of the sensing device cluster head node;

The sensing device cluster head node infers its defense strategy through a deep neural network. When an intelligent jamming attacker launches a multi-channel attack on the offload link, the game utility of maximizing the cluster head node of the sensing device can be formalized as follows:

In the multi-channel attack mode, a deep neural network defense model MSnet is established to maximize the game utility of the sensor equipment cluster head node. At the same time, MSnet is trained to infer the optimal defense strategy vector of the sensor device cluster head node in the multi-channel attack mode. The MSnet structure is shown in Figure 4.

The steps of MSnet processing the input and output of the channel gain of the cluster head node of the sensing device in the multi-channel attack mode are as follows:

①Input: The multi-channel gain vector of the sensor device node with a normalized mean value of 0 and a variance of 1

整形后变换成m个二维矩阵，分别对应于m个信道。② data shaping. go through

After shaping, it is transformed into m two-dimensional matrices, corresponding to m channels respectively.

(3) The shaped data is input to the convolution layer, and m convolution kernels with a stride of 1 are used to perform the first convolution operation to extract the multi-channel gain information of the key changes of the cluster head node of the sensing device.

其中

为权值。输出向量为

④ The single output of the convolutional layer is:

in

is the weight. The output vector is

为relu的单个输出，多输出向量为

⑤ The output of the convolutional layer uses relu as the activation function,

is the single output of relu, and the multi-output vector is

⑥ Then perform the second convolution operation, using m convolution kernels of 3×3 stride 1, and the single output of the convolution layer is:

多输出向量为

The multi-output vector is

为relu的单个输出，多输出向量为

⑦ Use relu as the activation function again,

is the single output of relu, and the multi-output vector is

多输出向量为

⑧In order to speed up the training speed, using the maximization pooling layer, after a 2×2 sliding window with a step size of 1, the single output of the maximization pooling layer is obtained as

The multi-output vector is

得到多输出向量为(P_M,1,P_M,2,...,P_M,m)。⑨ Finally, use the Sigmoid function to output

The multi-output vector is obtained as (P _M,1 ,P _M,2 ,...,P _M,m ).

The MSnet training and inference process is as follows:

初始化完成后，使用随机梯度下降法训练MSnet，逐步反向传播调整权值，使得传感设备簇头节点的博弈效用最大。其中，传感设备簇头节点通过交互获得智能干扰攻击者的传输功率。在多信道攻击模式下，传感设备簇头节点最大化其博弈效用的损失函数表示如下：The sensing device cluster head node infers its defense strategy through MSnet. Therefore, the cluster head node of the sensing device first randomly initializes the training weight vector

After the initialization is completed, MSnet is trained by stochastic gradient descent, and the weights are adjusted by back-propagation step by step, so that the game utility of the cluster head node of the sensing device is maximized. Among them, the cluster head node of the sensing device obtains the transmission power of the intelligent interference attacker through interaction. In the multi-channel attack mode, the loss function of the sensor device cluster head node to maximize its game utility is expressed as follows:

Among them, α _s,i represents the weight coefficient of the loss function, the influence of the balance constraint on the training process, (1-α _s,i )tanh(|P _i -P _max |) is the regularization term, the sensor device cluster head node The power constraints are involved in training the weights of MSnet.

By taking the partial derivative of the power in the loss function, we get:

The weight update equation for training MSnet is:

MSnet输出优化的功率策略向量通过随机梯度下降法，使得DNN处于收敛状态，基于DNN的Stackelberg博弈存在的纳什均衡点为(J_NN,P_MM)。where _θs is the learning rate. After the MSnet training is completed, use the trained MSnet to infer the transmission power policy vector of the sensing device cluster head node for defense. In the multi-channel attack mode, when inputting the channel gain vector of the cluster head node of the sensing device

MSnet outputs the optimized power policy vector through the stochastic gradient descent method, which makes the DNN in a convergent state. The Nash equilibrium point of the Stackelberg game based on DNN is (J _NN , P _MM ).

At this time, in the multi-channel attack mode, the utility of intelligently interfering with the attacker and the cluster head node of the sensing device reaches the maximum.

When m=n=1, that is, in the single-channel attack mode, the game utility of maximizing the cluster head node of the sensing device can be specially simplified as:

MSnet can be simplified as a single-layer SSnet, and SSnet can be trained to learn and reason the game strategy of the sensor device cluster head node in the single-channel attack mode, so as to maximize the game utility of the sensor device cluster head node.

The structure of SSnet is shown in Figure 5: the network model consists of normalization, full connection layer, data shaping, convolution layer, pooling layer, etc.

SSnet processes the input and output steps of the sensor device cluster head node channel gain as follows:

①In order to speed up the convergence speed of the game strategy learning of the sensing device cluster head node and ensure that the channel gain of the sensing device cluster head node is identically distributed. Therefore, the input channel gain is normalized to a standard normal distribution with mean 0 and variance 1. The normalized output of the channel gain of the sensor device cluster head node is:

输入SSnet全连接层。where E(·) is the expectation and D(·) is the variance. The normalized sensor device cluster head node channel gain vector

Enter the SSnet fully connected layer.

整形后变换成二维矩阵。② data shaping. The output of the fully connected layer goes through

After shaping, it is transformed into a two-dimensional matrix.

(3) The shaped data is input to the convolution layer, and the first convolution operation is performed using a 3×3 convolution kernel with a stride of 1, and the channel gain information of the key changes of the cluster head node of the sensing device is extracted.

其中

为权值。④ The output of convolutional layer 1 is:

in

is the weight.

为relu的输出。⑤ The output of convolutional layer one uses relu as the activation function,

is the output of relu.

⑥ Then perform the second convolution operation, and the output of the second convolution layer is:

为relu的输出。⑦ Use relu as the activation function again,

is the output of relu.

⑧ In order to speed up the training, use the maximization pooling layer and use a 2×2 sliding window with a stride of 1, and the output of the maximization pooling layer is

⑨ Using a fully connected layer, stretch the output of the maximizing pooling layer into an m×1 vector.

⑩ Finally, use the Sigmoid function to output

The SSnet training and inference process is as follows:

初始化完成后，使用随机梯度下降法训练SSnet，逐步反向传播调整权值，使得传感设备簇头节点的博弈效用最大。其中，传感设备簇头节点通过交互获得智能干扰攻击者的传输功率。传感设备簇头节点最大化其博弈效用的损失函数表示如下：The sensing device cluster head node infers its defense strategy through SSnet. Therefore, the cluster head node of the sensing device first randomly initializes the training weights

After the initialization is completed, the SSnet is trained using the stochastic gradient descent method, and the weights are adjusted by back-propagation step by step, so that the game utility of the cluster head node of the sensing device is maximized. Among them, the cluster head node of the sensing device obtains the transmission power of the intelligent interference attacker through interaction. The loss function of the sensor device cluster head node to maximize its game utility is expressed as follows:

Among them, α _s represents the weight coefficient of the loss function, the impact of the balance constraint on the training process, (1-α _s )tanh(|PP _max |) is the regularization term, and the power constraint of the sensor device cluster head node participates in the weight of training SSnet value.

By taking the partial derivative of the power P in the loss function, we get:

The weight update equation for training SSnet is:

SSnet输出优化的功率策略P_M。通过随机梯度下降法，使得DNN处于收敛状态，基于DNN的Stackelberg博弈存在的纳什均衡点为(J_N,P_M)。此时，在单信道攻击模式下，智能干扰攻击者和传感设备簇头节点的效用达到最大。where _θs is the learning rate. After the SSnet training is completed, use the trained SSnet to infer the transmission power strategy of the cluster head node of the sensing device for defense. When the channel gain vector of the cluster head node of the sensing device is input,

SSnet outputs an optimized power policy P _M . Through the stochastic gradient descent method, the DNN is in a state of convergence, and the Nash equilibrium point of the Stackelberg game based on DNN is (J _N , P _M ). At this time, in the single-channel attack mode, the utility of intelligently interfering with the attacker and the cluster head node of the sensing device reaches the maximum.

4) According to the power allocation strategy P _MM of the sensor device cluster head node obtained in step (3), the decision configuration variable is determined to perform task offloading.

In this embodiment, in the sensing device computing task offloading in the sensing edge cloud environment, a low-complexity and accurate defense is achieved when an intelligent interference attacker with learning ability attacks the sensing device computing task offloading link. Specifically through the following schemes:

In this embodiment, a computing task offloading capacity model of a sensor device cluster head node with sufficient resources is established for the sensing device computing task offloading scenario. A defense strategy is designed for the intelligent interference attacker with learning ability to attack the offloading process of computing tasks.

This embodiment specifically lists the optimization problem of maximizing the game utility of the intelligent jamming attacker and the sensor device cluster head node respectively under the single-channel attack mode. Establish a deep neural network model SJnet for intelligent jamming attackers to optimize game strategies in single-channel attack mode, and a deep neural network model SSnet for game defense strategy optimization by sensor device cluster head nodes, aiming to maximize the utility of game participants respectively. SJnet and SSnet are trained, and when the intelligent jamming attacker changes the single-channel attack strategy, the cluster head node of the sensing device uses SSnet inference in the single-channel attack mode to quickly obtain the optimal power allocation strategy to defend against intelligent jamming attacks.

Generally, in the multi-channel attack mode, the optimization problem of maximizing the game utility of the intelligent jamming attacker and the sensor device cluster head node is formulated respectively. Establish a deep neural network model MJnet for intelligent jamming attackers to optimize multi-channel attack game strategies in multi-channel attack mode, and a deep neural network model MSnet for sensor equipment cluster head nodes to optimize multi-channel defense strategies to maximize the multi-channel mode. The game utility of the intelligent jamming attacker and the sensor device cluster head node is trained to train MJnet and MSnet respectively, and when the intelligent jammer attacker changes the multi-channel attack strategy, the sensor device cluster head node uses MSnet in the multi-channel attack scenario Inference quickly obtains the optimal power allocation policy vector to defend against multi-channel attacks by intelligent jammers.

Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, etc., All should be included within the protection scope of the present invention.

Claims (8)

1. a game defense strategy optimization method of being attacked by intelligent interference in a sensing edge cloud, is characterized in that, comprises the following steps: (1) Obtain the transmission power vector P of the initial sensor device cluster head node set allocated to the computing task: P=(P ₁ , P ₂ ,...,P _m ), where m is the available sensor device cluster head node The number of channel resources; (2) According to the Stackelberg model, with the device cluster head node as the leader and the intelligent jamming attacker as the follower, according to the channel gain vector of the n sensing device cluster head nodes attacked by the intelligent jamming attacker, Calculate when the game utility of the intelligent jamming attacker is maximized, the power allocation vector J _NN of the intelligent jamming attacker to the n channels it attacks, as the power allocation strategy of the intelligent jamming attacker; The game utility is calculated as follows:

Among them, n is the total number of channels attacked by the intelligent jamming attacker, P is the transmission power vector of the sensor device cluster head node set allocated to the computing task, and J is the transmission power vector of the intelligent jamming attacker J=(J ₁ , _{J 2 , .} The ith channel of the node is used to unload the computing task, otherwise a _s,i =0; h _s,i is the computing task of the upstream computing task of the s th sensor device cluster head node on the ith channel offload link channel gain, P _i is the transmission power of the sensor device cluster head node of the i-th channel, n _0,i is the noise power of the i-th channel, h _J,i is the channel gain of the intelligent jamming attacker in the i-th channel, and J _i is The transmission power of the intelligent jamming attacker in the i-th channel, γ is the jamming attack cost per unit jamming power of the intelligent jamming attacker; (3) According to the Stackelberg model, on the premise that the intelligent jamming attacker adopts the power allocation strategy of step (2), according to the channel gain vectors of the m channels of the cluster head node of the sensing device, calculate the maximum When the game utility of the sensing device cluster head node reaches the Nash equilibrium point, the transmission power distribution vector of the sensing device cluster head node to m available channels is used as the power distribution strategy of the sensing device cluster head node; so The calculation method of the game utility of the cluster head node of the sensing device is as follows:

Among them, m is the total number of channel resources available to the sensor device cluster head node, P is the transmission power vector of the sensor device cluster head node set allocated to the computing task, and J is the transmission power vector of the intelligent jamming attacker J=(J ₁ , _{J 2 , .} The ith channel of the node is used to unload the computing task, otherwise a _s,i =0; h _s,i is the computing task of the upstream computing task of the s th sensor device cluster head node on the ith channel offload link channel gain, P _i is the transmission power of the sensor device cluster head node of the i-th channel, n _0,i is the noise power of the i-th channel, h _J,i is the channel gain of the intelligent jamming attacker in the i-th channel, and J _i is The transmission power of the intelligent jamming attacker in the i-th channel, and λ is the transmission cost per unit transmission power of the cluster head node of the sensing device; (4) According to the power allocation strategy P _MM of the sensor device cluster head node obtained in step (3), the decision configuration variable is determined to perform task offloading. 2. The game defense strategy optimization method of being attacked by intelligent interference in the sensing edge cloud as claimed in claim 1, wherein the calculation in step (2) maximizes the game utility of the intelligent interference attacker. The power allocation strategy of the intelligent jammer is to use the deep neural network to establish the intelligent attack model. According to the channel gain vector H _s,i =(h _s,1 ,h _s of the n sensor equipment cluster head nodes attacked by the intelligent jammer _{, 2} ,...,h _s,n ), predict the power allocation strategy of the intelligent jamming attacker. 3. the game defense strategy optimization method of being attacked by intelligent interference in the sensing edge cloud as claimed in claim 2, is characterized in that, the game utility of maximizing intelligent interference attacker described in step (2) is denoted as:

Among them, J _max is the maximum transmission power of the intelligent jamming attacker; The intelligent attack model established by the deep neural network includes an input layer, a normalization layer, a fully connected layer, a data shaping layer, a convolution module, a pooling layer group, and an output layer that are sequentially connected; The input layer is used to input the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,n of the n sensor device cluster head nodes attacked by the intelligent jamming attacker ) to the normalization layer; The normalization layer normalizes the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,n ) of the sensor device cluster head node to obtain a normalized The normalized sensor device cluster head node channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,n ), and input to the data shaping layer through the fully connected layer; The data shaping layer is used to transform the output of the fully connected layer into a two-dimensional matrix after reshape (h _{s, i} ), and input it to the convolution layer; The convolution module includes two groups of convolution layers, the two groups of convolution layers are respectively connected by the Relu linear rectification function, and each group of convolution layers includes n convolution kernels; the normalized sensor device cluster after shaping The channel gain of the head node, after the first convolution operation is performed through the corresponding convolution kernel, is output through the activation function, and the second convolution operation is performed through the corresponding convolution kernel again, and after the activation function, it is output to the pooling layer group; The pooling layer group includes n parallel pooling layers and fully connected layers; The output layer is used to output the intelligent jamming attacker transmission power vector (J _N,1 ,J _N,2 ,...,J _N,n ) of the Nash equilibrium point existing in the Stackelberg game, which is the intelligent jamming Attacker's power allocation strategy. 4. the game defense strategy optimization method of being attacked by intelligent interference in the sensing edge cloud as claimed in claim 3, it is characterized in that, described adopting the intelligent attack model of deep neural network to establish, according to the following method training acquisition: Randomly initialize the multi-channel training weight vector

The gradient descent method is used for training, and the weights are gradually adjusted by anti-phase propagation. The intelligent interference attacker obtains the transmission power of the cluster head node of the sensing device through interaction; the loss function of the intelligent attack model is expressed as follows:

Among them, α _J,i represents the weight coefficient of the loss function, (1-α _J,i )tanh(|J _i -J _max |) is the regularization term, and the power constraint of the intelligent interference attacker participates in the training,

The weight update equation for training the intelligent attack model is:

where _θJ represents the learning rate. 5. The game defense strategy optimization method of being attacked by intelligent interference in the sensing edge cloud as claimed in claim 1, wherein, when the calculation maximizes the game utility of the cluster head node of the sensing device so as to reach the Nash equilibrium point, For the transmission power distribution strategy of the sensor equipment cluster head node, a deep neural network is used to establish an intelligent defense model, and the power distribution strategy of the sensor equipment cluster head node is obtained. 6. The game defense strategy optimization method of being attacked by intelligent interference in the sensing edge cloud as claimed in claim 5, is characterized in that, the game utility of the described maximizing sensing device cluster head node in step (3) is denoted as:

Among them, _Pmax is the maximum transmission power of the sensor device cluster head node; The intelligent defense model established by using a deep neural network includes an input layer, a normalization layer, a fully connected layer, a data shaping layer, a convolution module, a pooling layer group, and an output layer that are sequentially connected; The input layer is used to input the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,m ) of the m channels of the sensor device cluster head node to normalize chemical layer; The normalization layer normalizes the channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,m ) of the sensor device cluster head node to obtain a normalized The normalized sensor device cluster head node channel gain vector H _s,i =(h _s,1 ,h _s,2 ,...,h _s,m ), and input to the data shaping layer through the fully connected layer; The data shaping layer is used to transform the output of the fully connected layer into a two-dimensional matrix after reshape (h _{s, i} ), and input it to the convolution layer; The convolution module includes two groups of convolution layers, the two groups of convolution layers are respectively connected by the Relu linear rectification function, and each group of convolution layers includes m convolution kernels; the normalized sensor device cluster after shaping The channel gain of the head node, after the first convolution operation is performed through the corresponding convolution kernel, is output through the activation function, and the second convolution operation is performed through the corresponding convolution kernel again, and after the activation function, it is output to the pooling layer group; The pooling layer group includes m parallel pooling layers and fully connected layers; The output layer is used to output the output power vector P _MM =(P _M,1 ,P _M,2 ,...,P _M,m ) of the sensor device cluster head node of the Nash equilibrium point existing in the Stackelberg game, That is, the power allocation strategy of the cluster head node of the sensor device. 7. The game defense strategy optimization method of being attacked by intelligent interference in the sensing edge cloud as claimed in claim 6, is characterized in that, the intelligent defense model described in step (3) is trained and obtained according to the following method: Randomly initialize the multi-channel training weight vector

The gradient descent method is used for training, and the weights are gradually adjusted by inverse propagation. The cluster head node of the sensing device obtains the transmission power of the intelligent interference attacker through interaction; the loss function of the intelligent attack model is expressed as follows:

Among them, α _s represents the weight coefficient of the loss function, which balances the influence of constraints on the training process; (1-α _s )tanh(|PP _max |) is the regularization term, and the power constraint of the sensor device cluster head node participates in the training,

By taking the partial derivative of the power in the loss function, we get:

The weight update equation for training the intelligent defense model is:

where _θs is the learning rate. 8. A game defense strategy optimization system attacked by intelligent interference in a sensing edge cloud, characterized in that it comprises an initialization module, an intelligent interference attacker prediction module, a defense strategy decision module, and a configuration module; The initialization module is used to obtain the initial sensor device cluster head node set assigned to the calculation task transmission power vector P: P=(P ₁ , P ₂ , . . . , P _m ) and submit it to the intelligent jamming attack is the user prediction module, where m is the number of available channel resources of the cluster head node of the sensing device; The intelligent jamming attacker prediction module is used to follow the Stackelberg model, with the device cluster head node as the leader and the intelligent jamming attacker as the follower, according to the n sensing devices attacked by the intelligent jamming attacker The channel gain vector of the channel of the cluster head node, when the game utility of the intelligent jamming attacker is maximized, the power allocation vector J _NN of the intelligent jamming attacker to the n channels it attacks is calculated as the power of the intelligent jamming attacker The allocation strategy is submitted to the defense strategy decision-making module; the calculation method of the game utility of the intelligent interference attacker is as follows:

Among them, n is the total number of channels attacked by the intelligent jamming attacker, P is the transmission power vector of the sensor device cluster head node set allocated to the computing task, and J is the transmission power vector of the intelligent jamming attacker J=(J ₁ , _{J 2 , .} The ith channel of the node is used to unload the computing task, otherwise a _s,i =0; h _s,i is the computing task of the upstream computing task of the s th sensor device cluster head node on the ith channel offload link channel gain, P _i is the transmission power of the sensor device cluster head node of the i-th channel, n _0,i is the noise power of the i-th channel, h _J,i is the channel gain of the intelligent jamming attacker in the i-th channel, and J _i is The transmission power of the intelligent jamming attacker in the i-th channel, γ is the jamming attack cost per unit jamming power of the intelligent jamming attacker; The defense strategy decision-making module is used to calculate the maximum value according to the channel gain vectors of the m channels of the cluster head node of the sensing device according to the Stackelberg model and on the premise that the intelligent jamming attacker adopts a power allocation strategy. When the game utility of the sensing device cluster head node is optimized to reach the Nash equilibrium point, the transmission power distribution vector of the sensing device cluster head node to m available channels is used as the power distribution strategy of the sensing device cluster head node, Submit to the configuration module; the calculation method of the game utility of the cluster head node of the sensing device is as follows:

Among them, m is the total number of channel resources available to the sensor device cluster head node, P is the transmission power vector of the sensor device cluster head node set allocated to the computing task, and J is the transmission power vector of the intelligent jamming attacker J=(J ₁ , _{J 2 , .} The ith channel of the node is used to unload the computing task, otherwise a _s,i =0; h _s,i is the computing task of the upstream computing task of the s th sensor device cluster head node on the ith channel offload link channel gain, P _i is the transmission power of the sensor device cluster head node of the i-th channel, n _0,i is the noise power of the i-th channel, h _J,i is the channel gain of the intelligent jamming attacker in the i-th channel, and J _i is The transmission power of the intelligent jamming attacker in the i-th channel, and λ is the transmission cost per unit transmission power of the cluster head node of the sensing device; The configuration module is configured to determine decision configuration variables according to the power distribution strategy P _MM of the sensor device cluster head node, and perform task offloading.

Images