CN108287989A - A kind of man-machine recognition methods of sliding identifying code based on track - Google Patents

Abstract

The man-machine recognition methods of sliding identifying code based on track that the invention discloses a kind of, includes the following steps：Acquire user trajectory data；Multidimensional characteristic system is built according to track data；Track differentiation is carried out to multidimensional characteristic system according to designed man-machine identification model.In the present invention; by combining two kinds of phenomenons of mankind track to carry out the design of multidimensional characteristic system; it is accustomed to the sliding identifying code of feature description user; and then user's operation and machine imitation are distinguished; advantage can be occupied in the confrontation with the black production tool of attacker, plays preferable confrontation protective effect.

Description

A Trajectory-Based Sliding Verification Code Human-Machine Recognition Method

technical field

The invention relates to the technical field of biometric authentication, in particular to a trajectory-based human-machine identification method for sliding verification codes.

Background technique

As a biometric authentication technology, the sliding verification code can meet the security requirements of the current network environment for identity authentication. It has been widely used in a variety of man-machine verification products. This verification method is not only easy for users to understand and remember, but also greatly increases the violence. Crack difficulty. At the same time, it has also attracted the attention of attackers. Attackers have developed black production tools that can imitate human behavior and began to challenge the mouse trajectory during the verification process of sliding verification codes.

The attackers generate human-like trajectory batch operations through black production tools to bypass detection, and continuously upgrade their forged data during the confrontation process to continue to bypass the same upgraded detection technology. Existing detection technologies are mainly aimed at machine identification, and the way of fighting against constantly updated machine behavior has a lag, and detection updates are often after certain losses are caused by black production tools. Therefore, in the escalating technological confrontation between both sides, how to gain an advantage in the confrontation with the attacker's black production tools is particularly important.

Contents of the invention

The purpose of the present invention is to provide a trajectory-based man-machine recognition method for sliding verification codes, aiming to build a multi-dimensional effective feature system to identify the triggerer of sliding verification codes and ensure the security of the network environment protected by the verification behavior.

The technical scheme that the present invention adopts is as follows:

A trajectory-based human-machine recognition method for sliding verification codes, comprising the following steps:

S1: collect user trajectory data;

S2: Construct a multi-dimensional feature system based on trajectory data;

S3: According to the designed human-machine recognition model, the trajectory of the multi-dimensional feature system is distinguished.

Further, the multi-dimensional feature system includes X features, Y features, and T features.

Further, the specific steps of the X feature extraction are as follows:

S201: Extracting the X feature class, and normalizing the horizontal coordinate x of the trajectory;

S202: Divide the horizontal coordinates of the trajectory into the first half and the second half;

S203: Extract multiple X feature groups of the first half of the trajectory x front, the second half x rear, the first half of the leading position difference x front diff, the second half of the leading position difference x rear diff, and the stop segment final stop;

S204: Extract features in each X feature group, including maximum value, peak value, median value, variance, minimum value, and extreme difference.

Further, the specific steps of the Y feature extraction are as follows:

S211: Extract the Y feature class, and perform normalization processing on the longitudinal coordinate y of the trajectory;

S212: Extract multiple Y feature groups of the entire trajectory y, the half y half, the entire adjacent difference y diff, and the adjacent difference y diff diff of the entire trajectory;

S213: Extract features in each Y feature group, including variance, average, range, and value.

Further, the specific steps of the T feature extraction are as follows:

S221: extract T feature class, and perform normalization processing on time feature t;

S222: Extract the T-X feature group, and subtract the normalized time feature t from the normalized horizontal coordinate x;

S223: Extract features in the T-X feature group, including maximum value, peak value, median value, variance, minimum value, and extreme difference.

Further, the specific steps of the design of the human-machine recognition model in step S3 are as follows:

S301: Input the features in the multi-dimensional feature system into multiple training models for algorithm training;

S302: Linearly weight the training output of the feature algorithm.

Further, the training model includes: CatBoost model, XGBoost model, RandomForest model, LogisticRegression model.

In summary, owing to adopting above-mentioned technical scheme, the beneficial effect of the present invention is:

1. In the present invention, the multi-dimensional feature system is designed by combining the two phenomena of human trajectories, using features to describe the user's sliding verification habits, and then distinguishing user operations from machine imitation. It can take advantage in the confrontation and play a better role in protection against confrontation.

2. In the present invention, by using the horizontal feature x as the main method, the behavior habits of "people" when performing sliding verification codes are described, and the vertical feature y is used to describe the characteristics of "machines", and the time feature t is used as a supplement to describe the relationship between "people" and The difference between "machines" can more accurately distinguish user operations from machine imitation, and improve the accuracy of trajectory distinction.

3. In the present invention, the actual verification effect is on a test set of 2 million track records, and the harmonic F value of accuracy and recall reaches 88.56, which is much higher than the effect of 87.89 for the scheme mainly describing "machines".

Description of drawings

Fig. 1 is a feature relationship diagram of the multi-dimensional feature system of the present invention;

Fig. 2 is a conceptual diagram of the multi-dimensional feature system of the present invention;

Fig. 3 is a relational diagram of the human-machine recognition model of the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

Example 1

A trajectory-based human-machine recognition method for sliding verification codes, comprising the following steps:

S1: collect user trajectory data;

Collect user trajectory data (x, y, t), including the horizontal coordinate x and vertical coordinate y at different time points t during the trajectory triggering process. Specifically, it is to obtain the trajectory record of the user during the sliding verification code trigger process, Provide data support for the construction of the multi-dimensional feature system of the sliding verification code code.

S2: Construct a multi-dimensional feature system based on trajectory data;

Based on the discovery of two modes, one is the end-retracement phenomenon of human trajectories; the other is the phenomenon of long-distance and short-distance human trajectories; multi-dimensional feature extraction is carried out, and then a multi-dimensional feature system is constructed to facilitate the algorithm training of the human-machine recognition model in the later stage. Use characteristics to describe "people" instead of using characteristics to describe "machines", and find the difference between the two.

S3: According to the designed man-machine recognition model, the trajectory of the multi-dimensional feature system is distinguished;

Input the multi-dimensional feature system features into the designed man-machine recognition model to optimize the man-machine recognition model, input the multi-dimensional feature body features into the man-machine recognition model for feature learning, obtain different probability values, and make the machine-machine recognition model through linear weighting The output probabilities are closer to the true types of trajectories.

In the present invention, through the analysis of the human trajectory, based on the discovery of two modes, the first mode: the turning back phenomenon of the human trajectory, and the second mode: the rapid and slow phenomenon of the human trajectory, the construction of a multi-dimensional feature system is carried out, and then the human As shown in Figure 1, the human-machine recognition model is trained by constructing a multi-dimensional feature system for the training set corresponding to the user trajectory, and the same multi-dimensional feature system is constructed for the prediction set corresponding to the predicted trajectory. Input the trained man-machine recognition model for model optimization and trajectory classification training, and linearly weight the training output to obtain the probability value of distinguishing trajectory categories.

Example 2

On the basis of Example 1, the multi-dimensional feature system includes X features, Y features, and T features.

The present invention mainly uses the horizontal feature x to describe the behavior habits of "people" when performing sliding verification, and uses the longitudinal feature y to describe the characteristics of "machines", and uses the time feature T as a supplement to describe the difference between "human" and "machine". as shown in picture 2.

Further, the specific steps of the X feature extraction are as follows:

S201: Extracting the X feature class, and normalizing the horizontal coordinate x of the trajectory;

S202: Combining the phenomenon of "far, fast, near slow" in human trajectory mode 2, the trajectory will be divided into the first half and the second half;

Specifically, "far, fast, near, slow" indicates that in the process of sliding the verification code, the speed is faster when the distance from the target point is farther, and the speed is slower when the distance is closer to the target point. Therefore, in the structure of the horizontal x feature group, the trajectory is divided into the front half and the second half, which are extracted separately.

S203: Extract multiple X feature groups of the first half of the trajectory x front, the second half x rear, the first half of the leading position difference x front diff, the second half of the leading position difference x rear diff, and the stop segment final stop;

Combined with the phenomenon of "terminal retracement" in human trajectory mode 1, the final stop feature group of the extraction stop segment is constructed.

Specifically, the horizontal coordinate data of the entire trajectory is extracted from the trajectory data (x,y,t) to form a horizontal sequence {x ₁ ,x ₂ ,...,x _t ,...,x _n }, and the trajectory The first half of the sequence {x ₁ ,x ₂ ,...,x _n/2 } forms x_front, and the second half of the trajectory sequence {x _n/2 ,x _n/2+1 ,...,x _n } form x_rear, take the adjacent position difference {x ₂ -x ₁ ,x ₃ -x ₂ ,...} in the first half of the trajectory sequence to form x_front_diff, take the adjacent position difference {. .., x _n-1 -x _n-2 , x _n -x _n-1 } form x_rear_diff, combined with human trajectory mode 1, take the last fifth of the trajectory sequence to form the stop segment final_stop.

S204: Extract features in each X feature group, including maximum value, peak value, median value, variance, minimum value, and extreme difference.

In the present invention, the lateral features of the feature system are designed from multiple dimensions, which can better provide input for the model.

Further, the specific steps for obtaining the Y feature are as follows:

S211: Extract the Y feature class, and perform normalization processing on the longitudinal coordinate y of the trajectory;

Specifically, the longitudinal coordinate data of the entire trajectory is extracted from the trajectory data (x,y,t) to form a horizontal sequence {y ₁ ,y ₂ ,...,y _t ,...,y _n }, for the sequence Perform normalization.

S212: Extract multiple Y feature groups of the entire trajectory y, the half y half, the entire adjacent difference y diff, and the adjacent difference y diff diff of the entire trajectory;

Specifically, for the normalized longitudinal sequence {y ₁ ,y ₂ ,...,y _t ,...,y _n }, take the entire segment {y ₁ ,y ₂ ,.. .,y _t ,...,y _n } form y, take the entire segment {y ₁ ,y ₂ ,...,y _t ,...,y _n } in the trajectory sequence and subtract 0.5 to form y_half , take the adjacent position difference {y ₂ -y ₁ ,y ₃ -y ₂ ,...} of the entire segment in the trajectory sequence to form y_diff, and take the adjacent position difference of the entire segment in the trajectory sequence {(y ₃ -y ₂ )-(y ₂ -y ₁ ),...} form y_diffdiff.

S213: Extract features in each Y feature group, including variance, average, range, and value.

When the horizontal feature X is weakly descriptive on a certain trajectory, the vertical feature y that is more descriptive of the "machine" can play an auxiliary role in model discrimination.

Further, the specific steps of obtaining the T feature are as follows:

S221: extract T feature class, and perform normalization processing on time feature t;

Specifically, the independent time series is just a sampling mark, which does not have good interpretability, but it has a better expressive meaning when combined with the horizontal x-series.

The normalized X minus T is used to represent the velocity during trajectory generation on another level, providing better feature input to the model.

S222: Extract the T-X feature group, and subtract the normalized time feature t from the normalized horizontal coordinate x;

Specifically, the horizontal coordinate data of the entire trajectory is extracted from the trajectory data (x,y,t) to form a horizontal sequence {x ₁ ,x ₂ ,...,x _t ,...,x _n } and a time series {t ₁ ,t ₂ ,...,t _n }, normalize respectively, and then subtract the time series from the horizontal series {x ₁ ,x ₂ ,...,x _t ,...,x _n } {t ₁ ,t ₂ ,...,t _n }.

S223: Extract features in the T-X feature group, including maximum value, peak value, median value, variance, minimum value, and extreme difference.

In summary, the feature groups and feature lists provided in this embodiment are shown in Table 1:

Example 3

On the basis of embodiment 1, the specific steps of the design of the human-machine recognition model of the step S3 are as follows:

S301: Input the features in the multi-dimensional feature system into multiple training models for algorithm training;

S302: Perform linear weighting on the training output of the feature algorithm;

Further, the training model includes: CatBoost model, XGBoost model, RandomForest model, LogisticRegression model.

As shown in Figure 3, specifically, the probability values of the training outputs of the CatBoost model, XGBoost model, RandomForest model, and LogisticRegression model are linearly weighted to obtain a human-machine recognition model linearly weighted by the four basic models.

The above descriptions 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 should be included in the protection of the present invention. within range.

Claims (7)

1. a track-based sliding verification code man-machine recognition method, is characterized in that, comprises the following steps: S1: collect user trajectory data; S2: Construct a multi-dimensional feature system based on trajectory data; S3: According to the designed human-machine recognition model, the trajectory of the multi-dimensional feature system is distinguished. 2. A track-based human-machine recognition method for sliding verification codes according to claim 1, wherein the multi-dimensional feature system includes X features, Y features, and T features. 3. according to claim 2 a kind of sliding verification code man-machine recognition method based on track, it is characterized in that, described X feature extracts specific steps as follows: S201: Extracting the X feature class, and normalizing the horizontal coordinate x of the trajectory; S202: Divide the horizontal coordinates of the trajectory into the first half and the second half; S203: Extract multiple X feature groups of the first half of the trajectory x front, the second half x rear, the first half of the leading position difference x front diff, the second half of the leading position difference x rear diff, and the stop segment final stop; S204: Extract features in each X feature group, including maximum value, peak value, median value, variance, minimum value, and extreme difference. 4. according to claim 2, a kind of sliding verification code man-machine recognition method based on track, it is characterized in that, described Y feature extraction specific steps are as follows: S211: Extract the Y feature class, and perform normalization processing on the longitudinal coordinate y of the trajectory; S212: Extract multiple Y feature groups of the entire trajectory y, the half-half y half, the entire adjacent difference y diff, and the adjacent difference ydiff diff of the entire adjacent difference; S213: Extract features in each Y feature group, including variance, average, range, and value. 5. a kind of trajectory-based man-machine recognition method for sliding verification codes according to claim 2, is characterized in that, the specific steps of said T feature extraction are as follows: S221: extract T feature class, and perform normalization processing on time feature t; S222: Extract the T-X feature group, and subtract the normalized time feature t from the normalized horizontal coordinate x; S223: Extract features in the T-X feature group, including maximum value, peak value, median value, variance, minimum value, and extreme difference. 6. according to claim 1, a kind of sliding verification code human-machine recognition method based on track, it is characterized in that, described step S3 human-machine recognition model design specific steps are as follows: S301: Input the features in the multi-dimensional feature system into multiple training models for algorithm training; S302: Linearly weight the training output of the feature algorithm. 7. A track-based human-machine recognition method for sliding verification codes according to claim 6, wherein the training models include: CatBoost models, XGBoost models, RandomForest models, and LogisticRegression models.