Breach point testing method and related device, electronic device, and storage medium

Abstract

The present application discloses a method for testing a breach point and related devices, electronic devices, and storage media, wherein the method for testing a breach point includes: obtaining test status information of a robot arm completing a target action at a first moment; wherein the test status information includes joint posture information of the robot arm; obtaining a liveness detection result of a liveness attack sample; wherein the liveness attack sample is set on the robot arm; based on the liveness detection result and the test status information, obtaining the breach point information of the liveness detection and the action to be performed by the robot arm at a second moment; wherein the second moment is the next moment of the first moment. The above scheme can improve the efficiency of the breach point test.

Description

Technical Field

[0001] The present application relates to the field of automation technology, and in particular to a breakthrough point testing method and related devices, electronic equipment, and storage media. Background Art

[0002] With the development of information technology, liveness detection has been applied in many fields such as mobile payment and identity authentication. At the same time, there are also liveness attack samples (such as face photos, masks, etc.) that attack liveness detection, which affects the security of mobile payment and identity authentication. Therefore, it is necessary to test the breach point of liveness detection to optimize liveness detection and improve its security.

[0003] At present, the existing breach point test mainly relies on manual work, for example, the breach point test that manually holds a live attack sample for liveness detection is inefficient. In view of this, how to improve the efficiency of the breach point test has become an urgent problem to be solved. Summary of the invention

[0004] The present application provides a breakthrough point testing method and related devices, electronic equipment, and storage medium.

[0005] The first aspect of the present application provides a breakthrough point testing method, comprising: obtaining test status information of a robotic arm completing a target action at a first moment; wherein the test status information includes joint posture information of the robotic arm; obtaining a liveness detection result of a liveness attack sample provided on the robotic arm; based on the liveness detection result and the test status information, obtaining breakthrough point information of the liveness detection and an action to be performed by the robotic arm at a second moment, wherein the second moment is the next moment of the first moment.

[0006] Therefore, the test state information of the robot arm completing the target action at the first moment is obtained, and the test state information includes the joint posture information of the robot arm, and the liveness detection result of the liveness attack sample is obtained, and the liveness attack sample is set on the robot arm, so that based on the liveness detection result and the test state information, the breakthrough point information of the liveness detection and the action to be performed by the robot arm at the second moment are obtained, so it is possible to automatically find the breakthrough point information by relying on the robot arm without manpower, so the efficiency of the breakthrough point test can be improved. In addition, since it is possible to predict the action to be performed at the second moment next to the first moment based on the liveness detection result and the test state information of the target action completed at the first moment, it is possible to avoid time-consuming traversal of all pre-designed actions during the breakthrough point test, and actively predict the action to be performed at the next moment based on the current liveness detection result and the test state information, so that the movement strategy of the robot arm is more targeted, which can be conducive to improving the speed of the breakthrough point test.

[0007] Among them, before obtaining the test status information of the robotic arm completing the target action at the first moment, it also includes: based on the target liveness detection result and target test status information of the robotic arm at the previous moment before the first moment, obtaining the target action of the robotic arm at the first moment, and controlling the robotic arm to complete the target action at the first moment.

[0008] Therefore, similar to the action to be performed by the robotic arm at the second moment, the target action completed by the robotic arm at the first moment is obtained by using the target liveness detection result and target test status information at the previous moment before the first moment, which can be beneficial for continuously looping execution and prediction of actions during the entire breakthrough point test process, thereby improving the efficiency and robustness of the breakthrough point test.

[0009] The detection result is obtained by detecting a test image of a live attack sample, and based on the live detection result and the test status information, the action to be performed by the robot arm at the second moment is obtained, including: when the live detection result of the live attack sample is non-live, according to the test status information and the test image, the action to be performed by the robot arm at the second moment is obtained; wherein the test image is obtained by shooting the live attack sample after the robot arm completes the target action.

[0010] Therefore, the detection result is set to be obtained by detecting the test image of the live attack sample, so that when the liveness detection result of the liveness attack sample is non-liveness, the action to be performed by the robot arm at the second moment is obtained according to the test state information and the test image, and the test image is obtained by shooting the liveness attack sample after the robot arm completes the target action, and then the action to be performed by the robot arm at the second moment can be obtained by combining the test state information and the real-world test image. Therefore, the error rate caused by predicting the action to be performed due to relying on a single information can be reduced, which can be beneficial to improving the accuracy of the action to be performed.

[0011] Among them, according to the test status information and the test image, the action to be executed by the robot arm at the second moment is obtained, including: feature extraction of the test status information to obtain a first feature, and feature extraction of the test image to obtain a second feature; fusing the first feature and the second feature to obtain a fused feature of the test status information and the test image; according to the fused feature, the action to be executed by the robot arm at the second moment is obtained.

[0012] Therefore, by performing feature extraction on the test status information to obtain the first feature, and performing feature extraction on the test image to obtain the second feature, the first feature and the second feature are fused to obtain the fused feature of the test status information and the test image, and then based on the fused feature, the action to be performed by the robot arm at the second moment is obtained. Therefore, the features of the test status information and the features of the real-world test image can be combined to obtain the action to be performed, which can help improve the accuracy of the action to be performed.

[0013] Among them, feature extraction is performed on the test status information to obtain a first feature, including: using a first extraction subnetwork of a first neural network to extract features from the test status information to obtain the first feature; feature extraction is performed on the test image to obtain a second feature, including: using a second extraction subnetwork of the first neural network to extract features from the test image to obtain the second feature; and fusion features are used to predict to obtain the action to be performed by the robot arm at the second moment, including: using the prediction subnetwork of the first neural network to predict the fusion features to obtain the action to be performed by the robot arm at the second moment.

[0014] Therefore, the first extraction subnetwork of the first neural network is used to extract features of the test state information to obtain the first feature, and the second extraction subnetwork of the first neural network is used to extract features of the test image to obtain the second feature, which can help improve the efficiency of feature extraction, and the prediction subnetwork of the first neural network is used to predict the fusion features to obtain the action to be performed by the robot arm at the second moment. Therefore, the first neural network can be used to realize feature extraction and action prediction, which can help improve the efficiency of the breakthrough point test.

[0015] Among them, the prediction sub-network of the first neural network is used to predict the fusion features to obtain the action to be executed by the robot arm at the second moment, including: using the prediction sub-network to process the fusion features to obtain the prediction probability values ​​of multiple candidate actions; and taking the candidate action corresponding to the largest prediction probability value as the action to be executed by the robot arm at the second moment.

[0016] Therefore, by using the prediction sub-network to process the fusion features, the predicted probability values ​​of multiple candidate actions are obtained, and the candidate action corresponding to the largest predicted probability value is used as the action to be executed by the robot arm at the second moment, which can help improve the success rate of the breakthrough point test.

[0017] Among them, the test status information also includes ambient lighting information, and the first extraction subnetwork of the first neural network is used to extract features of the test status information to obtain the first feature, including: using the first extraction subnetwork to perform matrix multiplication on the joint posture information and the ambient lighting information to obtain the first feature.

[0018] Therefore, when the test state information also includes ambient lighting information, the first extraction subnetwork is used to perform matrix multiplication on the joint posture information and the ambient lighting information to obtain the first feature, which can be helpful in fusing the ambient lighting information and the joint posture information, thereby helping to improve the feature richness of the first feature, and further helping to improve the accuracy of subsequent action predictions and improve the success rate of breakthrough point testing.

[0019] The action to be performed by the robot arm at the second moment is predicted by the first neural network, and the method also includes: obtaining test state information when the robot arm has not performed the action to be performed as the first state information, and obtaining the test state information after the robot arm completes the action to be performed as the second state information; and, using the liveness detection result, obtaining the optimized excitation value of the first neural network; wherein, when the liveness attack sample is detected as non-liveness, the optimized excitation value is a non-positive value, and when the liveness attack sample is detected as liveness, the optimized excitation value is a positive value, and the liveness detection result includes a detection probability value, and the optimized excitation value is positively correlated with the detection probability value; the first state information, the action to be performed, the optimized excitation value and the second state information are used as training samples to train the first neural network.

[0020] Therefore, by obtaining the test state information when the robot arm does not perform the action to be performed as the first state information, and using the test state information after the robot arm completes the action to be performed as the second state information, and using the liveness detection result, the optimized excitation value of the first neural network is obtained, and when the liveness attack sample is detected as non-live, the optimized excitation value is a non-positive value, and when the liveness attack sample is detected as live, the optimized excitation value is a positive value, and the liveness detection result includes a detection probability value, and the optimized excitation value and the detection probability value are positively correlated, so that the first state information, the action to be performed, the optimized excitation value and the second state information are used as training samples to train the first neural network, which can be beneficial to the first neural network in the process of predicting the action to be performed at the second moment, and because when the liveness attack sample is detected as non-live, the optimized excitation value is a non-positive value, and when the liveness attack sample is detected as live, the optimized excitation value is a positive value, it can be beneficial to enable the first neural network to learn the strategy of finding the breakthrough point more quickly according to the positive and negative and size of the optimized excitation value, which is beneficial to improve the speed of the breakthrough point test.

[0021] Among them, using the liveness detection result to obtain the optimized excitation value of the first neural network includes: determining that the liveness attack sample is detected as live based on the detection probability value, and determining the optimized excitation value as the product of the detection probability value and the first value and the sum of the second value; and / or determining that the liveness attack sample is detected as non-live based on the detection probability, and determining the optimized excitation value as the difference between the product of the detection probability value and the first value and the second value; wherein the first value and the second value are both positive numbers, and the first value is not greater than the second value.

[0022] Therefore, when the live attack sample is detected as a live body based on the detection probability value, the optimized incentive value is determined as the product of the detection probability value and the first value and the sum of the second value. When the live attack sample is detected as a non-live body based on the detection probability, the optimized incentive value is determined as the difference between the product of the detection probability value and the first value and the second value. The first value and the second value are both positive numbers, and the first value is not greater than the second value. Therefore, when the live attack sample is detected as a non-live body based on the detection probability, the optimized incentive value is calculated as a non-positive value, thereby stimulating the first neural network, which can reduce the probability of falling into a non-optimal solution by performing actions back and forth at an unbroken position, and further improve the speed of the breakthrough point test.

[0023] Among them, based on the liveness detection result and the test status information, the action to be executed by the robot arm at the second moment is obtained, including: when the liveness detection result of the liveness attack sample is live, the joint posture of the robot arm is reset to a preset initial value, and the preset initial action is obtained as the action to be executed by the robot arm at the second moment.

[0024] Therefore, when the liveness detection result of the liveness attack sample is liveness, by resetting the joint posture of the robotic arm to the preset initial value and obtaining the preset initial action as the action to be executed by the robotic arm at the second moment, the robotic arm can start a new round of breakthrough point testing, which can help increase the amount of breakthrough point information obtained from the test.

[0025] The method of obtaining the breach point information of the liveness detection based on the liveness detection result and the test status information includes: when the liveness detection result of the liveness attack sample is liveness, determining the test status information as the breach point information.

[0026] Therefore, when the live attack sample is detected as live, determining the test status information as the breach point information can be helpful for automatically saving the breach point information in a timely manner during the breach point test process, thereby facilitating subsequent optimization and improvement of liveness detection.

[0027] The liveness detection result is obtained by using the second neural network detection, and the breakthrough point testing method may also include: optimizing the second neural network using the breakthrough point information.

[0028] Therefore, optimizing the second neural network used for liveness detection through the breakthrough point information obtained from the breakthrough point test can help improve the performance of the second neural network.

[0029] The second aspect of the present application provides a breakthrough point testing device, including: a state information acquisition module, a detection result acquisition module and a prediction module, the state information acquisition module is used to obtain test state information of a robotic arm completing a target action at a first moment; wherein the test state information includes joint posture information of the robotic arm; the detection result acquisition module is used to obtain a liveness detection result of a liveness attack sample, and the liveness attack sample is provided on the robotic arm; the prediction module is used to obtain breakthrough point information of liveness detection and an action to be performed by the robotic arm at a second moment based on the liveness detection result and the test state information, wherein the second moment is the next moment of the first moment.

[0030] A third aspect of the present application provides an electronic device, comprising a memory and a processor coupled to each other, wherein the processor is used to execute program instructions stored in the memory to implement the breach point testing method in the first aspect.

[0031] A fourth aspect of the present application provides a computer-readable storage medium having program instructions stored thereon, which implement the breach point testing method in the first aspect when the program instructions are executed by a processor.

[0032] The above scheme obtains the test state information of the robot arm completing the target action at the first moment, and the test state information includes the joint posture information of the robot arm, and obtains the liveness detection result of the liveness attack sample, and the liveness attack sample is set on the robot arm, so as to obtain the breakthrough point information of the liveness detection and the action to be performed by the robot arm at the second moment based on the liveness detection result and the test state information, so it is possible to automatically find the breakthrough point information by relying on the robot arm without manpower, so it is possible to improve the efficiency of the breakthrough point test. In addition, because it is possible to predict the action to be performed at the second moment next to the first moment based on the liveness detection result and the test state information of the target action completed at the first moment, it is possible to avoid time-consuming traversal of all pre-designed actions during the breakthrough point test, and actively predict the action to be performed at the next moment based on the current liveness detection result and the test state information, so that the movement strategy of the robot arm is more targeted, which can be conducive to improving the speed of the breakthrough point test. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flowchart of an embodiment of the breakthrough point testing method of the present application;

[0034] Figure 2 yes Figure 1 A schematic diagram of a flow chart of an embodiment of step S13;

[0035] Figure 3 is a schematic diagram of a framework of a first neural network embodiment;

[0036] Figure 4 It is a flowchart of another embodiment of the breakthrough point test of the present application;

[0037] Figure 5 This is a schematic diagram of the state of an embodiment of the breakthrough point test of the present application;

[0038] Figure 6 It is a schematic diagram of the framework of an embodiment of a breakthrough point testing device of the present application;

[0039] Figure 7 It is a schematic diagram of the framework of an embodiment of the electronic device of the present application;

[0040] Figure 8 It is a schematic diagram of a framework of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0041] The scheme of the embodiment of the present application is described in detail below in conjunction with the drawings of the specification.

[0042] In the following description, for the purpose of explanation rather than limitation, specific details such as specific system structures, interfaces, and technologies are provided to facilitate a thorough understanding of the present application.

[0043] The terms "system" and "network" are often used interchangeably in this article. The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship. In addition, "many" in this article means two or more than two.

[0044] See also Figure 1 , Figure 1 This is a flow chart of an embodiment of the method for testing a breach point of the present application. Specifically, it may include the following steps:

[0045] Step S11: Acquire test status information of the robot arm completing the target action at the first moment.

[0046] In the disclosed embodiment, the test state information may specifically include the joint pose information of the robot arm. The robot arm may be specifically set to: a five-axis joint robot arm, a six-axis joint robot arm, etc. according to the actual application needs, which is not limited here. The joint pose information of the robot arm may specifically include the pose of the joint of the robot arm relative to the reference position in the robot arm, wherein the reference position may refer to a component at a fixed position in the robot arm, such as a robot arm base, and then the joint pose information of the robot arm may include the pose of the joint of the robot arm relative to the robot arm base. For example, when the robot arm is a six-axis joint robot arm, the joint pose information may include the poses of the six joints relative to the base, which may be specifically represented by a 6*1 vector; or, when the robot arm is a five-axis joint robot arm, the joint pose information may include the poses of the five joints relative to the base, which may be specifically represented by a 5*1 vector, and other scenarios may be deduced by analogy, and examples are not given one by one here. Alternatively, the reference position may also refer to the position of an object whose relative position relationship with a component at a fixed position in the robot arm will not change.

[0047] In an implementation scenario, in order to simulate the real scene of liveness detection and improve the accuracy of the breakthrough point test, the test status information may also include ambient lighting information. Specifically, in order to further improve the authenticity of the environmental simulation, a light source may be set in the south, north, east, west, northeast, southeast, southwest, and northwest of the robot arm, respectively, to simulate the ambient lighting in different scenes. At this time, the ambient lighting information can be represented by an 8*1 vector, and other scenes can be deduced by analogy, and examples are not given one by one here. It should be noted that in actual application, fewer or more light sources than the above examples may be deployed, and the specific adjustments can be made according to actual test needs.

[0048] It should be noted that the target action can be the initial action when starting the breakthrough point test, that is, the action of the robot arm when it moves for the first time in the complete process of testing to obtain a breakthrough point information, that is, the first moment can be the moment of starting the breakthrough point test; in addition, the target action can also be the action of the robot arm when it moves next time after completing the initial action, that is, the first moment can also be the moment from the start of the breakthrough point test to the process of obtaining the breakthrough point information.

[0049] In an implementation scenario, when the breakthrough point test is started, the test state information can be set to the initial state information. Specifically, the joint posture of the robot arm can be set to a preset initial value. The preset initial value can be set according to the actual application needs. For example, the preset initial value of the joint posture can be the default posture when the robot arm is powered on, or the posture before the robot arm was powered off the last time, etc., which is not limited here. In addition, when the initial state information includes ambient light information, the ambient light can also be set to another preset initial value. The preset initial value can also be set according to the actual application needs. For example, the preset initial value of the ambient light can be that all light sources are turned on, or all light sources are turned off, or some light sources are turned on and some light sources are turned off, etc., which is not limited here. It should be noted that in the process of deploying the ambient light, the light intensity, light angle, etc. of each light source in the open state can be adaptively adjusted in combination with the actual test needs. In addition, when starting the breakthrough point test, an initial action can also be determined so that the robot arm can start the breakthrough point test process according to the initial action.

[0050] In a specific implementation scenario, one can be randomly selected from the preset action set or specified according to certain rules as the initial action, and the initial action is used as the action to be executed after starting the breakthrough point test. The preset action set can be set according to actual application needs, for example, it can be set to include: action 1, action 2, action 3, ..., action N, and specific actions are not given here.

[0051] In another specific implementation scenario, a first neural network can also be pre-trained so that the initial state information can be used as input data of the first neural network, and prediction can be performed through the first neural network to determine the initial action. Specifically, the first neural network predicts the initial state information and can obtain the predicted probability value of each candidate action among multiple candidate actions, so that the candidate action corresponding to the maximum predicted probability value can be used as the initial action. In addition, in order to improve the prediction accuracy, it is also possible to obtain a test image obtained by shooting a live attack sample under the initial state information, so that the initial state information and the test image can be used as input data of the first neural network, and prediction can be performed through the first neural network to determine the initial action, and then the initial state information and the test image captured in the real world can be combined to jointly predict the action to be performed at the next moment, which is conducive to improving the prediction accuracy.

[0052] In another implementation scenario, during the breakthrough point test after the initial action is completed, the target action of the robotic arm at the first moment can be obtained based on the target liveness detection result and target test status information of the robotic arm at the previous moment before the first moment. Specifically, the target test status information is the test status information of the robotic arm after the action is completed at the previous moment. For details about the test status information, please refer to the above description, which will not be repeated here. In addition, the target liveness detection result at the previous moment may include whether the live attack sample set on the robotic arm at the previous moment is detected as live or non-live. For details about the liveness detection result, please refer to the following description, which will not be repeated here. Through the above method, it can be beneficial to continuously loop the current action and predict the next action during the entire breakthrough point test process, and the robotic arm can be controlled to actively execute according to the predicted next action to continue to complete the breakthrough point detection process, thereby helping to improve the efficiency and robustness of the breakthrough point test.

[0053] Step S12: Obtaining the liveness detection result of the liveness attack sample.

[0054] In the disclosed embodiment, the live attack sample can be arranged on the mechanical arm. Specifically, the live attack sample can be arranged at the end of the mechanical arm away from the base thereof.

[0055] In an implementation scenario, a liveness detection device may be used to perform liveness detection on liveness attack samples to obtain liveness detection results. Specifically, a test image obtained by shooting a liveness attack sample may be detected to obtain a liveness detection result. The liveness detection device may be provided with a camera, may have corresponding computing capabilities, or may have the ability to transmit data between other servers, clouds, etc. that have corresponding computing capabilities. This means that, in the implementation process, the liveness detection device may independently implement the liveness detection function, obtain corresponding liveness detection results, or use the capabilities of other devices to complete the liveness detection function. For example, a liveness detection device may include but is not limited to: a mobile phone, a tablet computer, a laptop computer, etc., which are not limited here. Liveness attack samples may include but are not limited to: an electronic portrait displayed on an electronic device, a printed paper portrait, a mask generated using a mold, a three-dimensional portrait such as a head mold, etc., which are not limited here.

[0056] In a specific implementation scenario, the liveness detection device can be set on the side of the robotic arm close to the liveness attack sample. Specifically, in order to improve the accuracy of liveness detection, the liveness detection device can be set in a position to clearly capture the image of the liveness attack sample. For example, it can be set in front of the robotic arm.

[0057] In another specific implementation scenario, the liveness detection device can run a neural network, specifically an artificial intelligence neural network, for liveness detection; alternatively, the liveness detection device can also be connected to a server, a cloud, etc. in communication, and the server, the cloud, etc. can run an artificial intelligence neural network for liveness detection, so that the liveness detection device can send images of liveness attack samples captured to the server, the cloud, etc., so that after the server, the cloud, etc. receives the corresponding data, liveness detection is performed and the liveness detection result is obtained.

[0058] In another specific implementation scenario, after the robotic arm completes the target action at the first moment, at least one of the joint posture of the robotic arm and the ambient light changes, so that the live attack sample set on the robotic arm can be photographed at this time to obtain a test image, and the test image is subjected to liveness detection to obtain a liveness detection result. Specifically, the liveness detection result may include a detection probability value. When the detection probability value is less than or equal to a preset probability threshold, the liveness detection result of the liveness attack sample can be considered to be live (i.e., under the test state information, the liveness detection device misjudges the liveness attack sample as live, i.e., successfully attacks the liveness detection). When the detection probability value is greater than the preset probability threshold, the liveness detection result of the liveness attack sample can be considered to be non-live (i.e., under the test state information, the liveness detection device successfully determines that the liveness attack sample is not live). Specifically, the preset probability threshold can be set according to actual application needs. For example, the size of the preset probability value is usually negatively correlated with the accuracy requirement of liveness detection. For example, when the accuracy requirement of liveness detection is relatively high, a lower preset probability threshold can be set, such as 0.2, 0.25, etc.; or, when the accuracy requirement of liveness detection is relatively low, a higher preset probability threshold can be set, such as 0.3, 0.35, etc., which is not limited here. Of course, the parameters that affect the preset probability value may include but are not limited to the accuracy requirement of liveness detection, and the specific setting method is not limited to the above-mentioned examples.

[0059] Step S13: Based on the liveness detection result and the test status information, the breakthrough point information of the liveness detection and the action to be performed by the robot arm at the second moment are obtained, wherein the second moment is the next moment of the first moment.

[0060] In one implementation scenario, the detection result is obtained by detecting a test image of a live attack sample, and the test image is obtained by taking the live attack sample after the robotic arm completes the target action at the first moment. Please refer to the above description for details, so that the action to be performed by the robotic arm at the second moment can be obtained based on the test status information and test image after the robotic arm completes the target action at the first moment.

[0061] In a specific implementation scenario, the first neural network can be used to predict the above-mentioned test state information and test image to obtain the action to be performed by the robotic arm at the second moment, which can help improve the prediction efficiency.

[0062] In another specific implementation scenario, the liveness detection result of the liveness attack sample may be live, that is, the liveness attack sample successfully breaks the liveness detection, or it may be non-live, that is, the liveness attack sample fails to successfully break the liveness detection. In the case that the liveness detection result of the liveness attack sample is non-live (that is, it fails to break through successfully), the action to be performed by the robot arm at the second moment can be obtained according to the test status information and the test image, so that the second moment can be used as the new first moment, and the action to be performed at the second moment can be used as the target action of the new first moment, and the above-mentioned steps of obtaining the test status information of the robot arm completing the target action at the first moment and subsequent steps are re-executed to automatically perform the breakthrough point test and improve the efficiency of the breakthrough point test.

[0063] In addition, when the liveness detection result of the liveness attack sample is liveness, the joint posture of the robot arm can also be reset to a preset initial value, and the preset initial action can be obtained as the action to be executed at the second moment of the robot arm, so that the second moment can be used as the new first moment, and the action to be executed at the second moment obtained is used as the target action of the new first moment, and the above-mentioned step of obtaining the test state information of the robot arm completing the target action at the first moment and the subsequent steps are re-executed, so as to continue the breakthrough point test when the liveness detection is successfully broken, so as to help increase the number of breakthrough points found. Specifically, during the breakthrough point test process, in order to obtain different breakthrough point information as much as possible, the preset initial action can be set to different in different test stages. For example, when the liveness detection result of a liveness attack sample is liveness, the preset initial action can be action 1, and when the liveness detection result of another liveness attack sample is liveness, the preset initial action can be action 2. The specific method of obtaining the preset initial action can refer to the above description, which will not be repeated here.

[0064] In an implementation scenario, when the liveness detection result of the liveness attack sample is liveness, the test state information can be determined as the breach point information. Specifically, when the liveness detection result after the robot arm completes the target action at the first moment indicates that the liveness attack sample is liveness, the test state information of the robot arm completing the target action at the first moment can be used as the breach point information. For example, the joint posture information, environmental lighting information, etc. contained in the test state information can be used as the breach point information.

[0065] In one implementation scenario, the liveness detection result is obtained by using the second neural network. Specifically, the test image of the liveness attack sample can be used as the input data of the second neural network, so that the test image is detected by the second neural network to obtain the liveness detection result. After obtaining the breach point information, the breach point information can also be used to optimize the second neural network. Specifically, the breach point information is the test state information of the robotic arm when the liveness attack sample is detected as a live body. The breach point information can be used to obtain the test image obtained by shooting the liveness attack sample when breaking the liveness detection, and the test image can be used as a training image marked with the actual category of "non-liveness", so that the second neural network can be used to classify the training image to obtain the predicted category of the training image (such as "liveness" and "non-liveness"), and then the network parameters of the second neural network can be adjusted according to the difference between the actual category and the predicted category. For example, the cross entropy loss function can be used to calculate the loss value of the actual category and the predicted category, and gradient optimization can be performed. In the above manner, the breach point information obtained by the test can be used to optimize the liveness detection performance of the second neural network and improve the liveness detection accuracy of the second neural network. Therefore, in application scenarios such as supermarkets and online shopping that require facial payment, application scenarios such as social security registration / login, brokerage account opening / login, etc. that require identity authentication, as well as applications (APPs) such as traffic management and payment that require identity authentication, the steps in the embodiments of the present disclosure can be used to obtain breakthrough point information, and then the breakthrough point information can be used to optimize the neural network used for liveness detection, so as to improve the accuracy of liveness detection as much as possible and reduce the probability of liveness detection being broken, thereby improving the security of facial payment, or improving the security of liveness verification in application scenarios such as social security and brokerages.

[0066] In an implementation scenario, when the first neural network is used to predict the action of the robot arm, and the second neural network is used to detect whether the live attack sample is live, the first neural network predicts the action to be performed by the robot arm at the next moment after the first moment (i.e., the second moment in the embodiment of the present disclosure) based on the test state information of the robot arm completing the target action at the first moment and the liveness detection result of the liveness attack sample by the second neural network, so that the second moment can be used as the new first moment, and the action to be performed at the second moment can be used as the target action of the new first moment, and this cycle is repeated until the liveness detection result of the liveness attack sample by the second neural network is live, indicating that the second neural network is successfully broken, so that the test state information of the robot arm when the second neural network is successfully broken can be used as the breaking point information. By repeatedly executing multiple rounds of the above-mentioned breaking point test process, several breaking point information of the second neural network can be obtained, so that the second neural network can be optimized by using several breaking point information in the above-mentioned manner to minimize the vulnerabilities of the second neural network and improve the security of liveness detection.

[0067] The above scheme obtains the test status information of the robot arm completing the target action at the first moment, and the test status information includes the joint posture information of the robot arm, and obtains the liveness detection result of the liveness attack sample, and the liveness attack sample is set on the robot arm, so as to obtain the breakthrough point information of the liveness detection and the action to be performed by the robot arm at the second moment based on the liveness detection result and the test status information, so the breakthrough point information can be automatically found by the robot arm, so the efficiency of the breakthrough point test can be improved. In addition, since the action to be performed at the second moment after the first moment can be predicted based on the liveness detection result and the test status information of the target action completed at the first moment, it is possible to avoid time-consuming traversal of all pre-designed actions during the breakthrough point test, and actively predict the action to be performed at the next moment based on the current liveness detection result and the test status information, so that the movement strategy of the robot arm is more targeted, which can be conducive to improving the speed of the breakthrough point test.

[0068] See also Figure 2 , Figure 2 yes Figure 1 Schematic diagram of the process of an embodiment of step S13 in FIG. Figure 2 The process is to use the predicted state information and the test image to perform prediction processing to obtain a flow chart of the action to be performed by the robot arm at the second moment. Specifically, the following steps may be included:

[0069] Step S21: extracting features from the test state information to obtain a first feature, and extracting features from the test image to obtain a second feature.

[0070] In an implementation scenario, when the test state information includes the joint posture information and the ambient light information of the robot arm, feature extraction can be performed on the joint posture information and the ambient light information to obtain a first feature.

[0071] In a specific implementation scenario, the joint posture information can be represented by a matrix, for example, a 6*1 vector, and the ambient light information can also be represented by a matrix, for example, an 8*1 vector. For details, please refer to the relevant description in the aforementioned disclosed embodiment, which will not be repeated here. Then, the above matrix can be feature extracted to obtain the first feature.

[0072] In another specific implementation scenario, in order to improve the efficiency of feature extraction, the first extraction sub-network of the first neural network can also be used to extract features from the test state information to obtain the first feature. Figure 3 , Figure 3 Schematic diagram of the framework of the first neural network embodiment. Figure 3As shown, the first extraction sub-network is used to perform matrix multiplication on the joint pose information and the ambient light information to obtain the first feature. Specifically, when the joint pose information is a 6*1 vector and the ambient light information is an 8*1 vector, the first extraction sub-network can be used to perform matrix multiplication on the joint pose information and the ambient light information to obtain a 64*1 vector, and the vector is preset so that the first feature and the second feature have the same size and channel. Please refer to Figure 3 When the second feature is 4*4*64, the 64*1 vector obtained by matrix multiplication can be channel-copied to obtain a second feature of 4*4*64.

[0073] In one implementation scenario, in order to improve efficiency, the second extraction subnetwork of the first neural network can be used to extract features from the test image to obtain the second feature.

[0074] In a specific implementation scenario, the second extraction subnetwork may specifically include at least one sequentially connected convolutional layer, for example, may include one convolutional layer, two convolutional layers, three convolutional layers, etc., which is not limited here.

[0075] In another specific implementation scenario, please refer to Figure 3 , the test image is a 128*28*3 image. The second extraction subnetwork includes two sequentially connected convolutional layers, which are respectively denoted as conv1 and conv2 for the convenience of description. Figure 3 As shown, conv1 includes 32 filters, each filter has a size of 8*8, and the convolution stride is 4. After conv1, the test image can extract a 16*16*32 feature map, and conv2 includes 64 filters, each filter has a size of 6*6, and the convolution stride is 2. After conv2, a 4*4*64 second feature can be extracted. In other implementation scenarios, the second extraction subnetwork can also be other network structures, which can be set according to actual application needs and are not limited here.

[0076] In the above method, the test state information can be considered as the information of the robot state space, and the test image can be considered as the information of the robot observation space. By extracting the features of the test state information, the robot state space can be encoded, and by extracting the features of the test image, the robot observation space can be encoded. Therefore, encoding can be performed in multiple dimensions, thereby improving feature richness.

[0077] Step S22: The first feature and the second feature are fused to obtain fused features of the test state information and the test image.

[0078] In the embodiments of the present disclosure, specific methods of fusion processing may include, but are not limited to, splicing, stacking, etc., which may be configured according to actual application needs and are not limited here.

[0079] In an implementation scenario, the first feature and the second feature have the same size and channel. In order to reduce the complexity of the fusion processing and increase the speed of the fusion processing, the first feature and the second feature can be directly spliced ​​to obtain the fusion features of the test status information and the test image.

[0080] In a specific implementation scenario, please refer to Figure 3 , the first feature is a 4*4*64 feature map, and the second feature is a 4*4*64 feature map, then the two can be directly spliced ​​to obtain a 4*4*128 feature map as a fusion feature. Other scenarios can be set according to actual application conditions, and no examples are given here.

[0081] In the above method, as mentioned above, the first feature can be considered as the coding information of the robot state space, and the second feature can be considered as the coding information of the robot observation space. Therefore, the fusion of the first feature and the second feature can fuse the robot state space and the observation space.

[0082] Step S23: Use the fusion features to make predictions to obtain the action to be performed by the robot arm at the second moment.

[0083] In an implementation scenario, in order to improve prediction efficiency, the prediction sub-network of the first neural network can be used to predict the fusion features to obtain the action to be performed by the robotic arm at the second moment.

[0084] In a specific implementation scenario, the prediction subnetwork is used to process the fusion features to obtain the predicted probability values ​​of multiple candidate actions, so that the candidate action corresponding to the maximum predicted probability value can be used as the action to be executed by the robot at the second moment. Specifically, the multiple candidate actions can include 100 actions, 1,000 actions, 10,000 actions, etc., which can be set according to actual application needs and are not limited here.

[0085] In another specific implementation scenario, the prediction subnetwork may include sequentially connected convolutional layers and fully connected layers. Figure 3 , for the convenience of description and to distinguish from the convolutional layers in the second extraction sub-network, the convolutional layers of the prediction sub-network are denoted as conv3 and conv4, and the fully connected layers are denoted as fc1 and fc2. Figure 3As shown, conv3 includes 64 filters, each of which has a size of 4*4 and a convolution step of 1, so that the 4*4*128 fusion feature can be extracted to obtain a 4*4*64 feature map, and conv4 includes 64 filters, each of which has a size of 3*3 and a convolution step of 1, so that the 4*4*64 feature map can be further extracted to obtain a 2*2*64 feature map, and a 256*1 feature map can be obtained through dimension conversion, and after two layers of fully connected layers with input and output of 256, an output vector of 256*1 is obtained, and the action to be executed at the second moment can be obtained through the output vector. In addition, the vector is subjected to matrix multiplication, and the test state information that is most likely to break the liveness detection at the second moment can also be obtained. Specifically, it can include joint posture information and ambient light information. For the convenience of description, the test state information that is most likely to break the liveness detection at the second moment is recorded as s′. In the embodiment of the present application, unless otherwise specified, s′ is used to represent the test state information that is most likely to break the liveness detection at the second moment predicted by the first neural network. In addition, in other scenarios, the relevant parameters of each convolutional layer and fully connected layer in the first neural network can also be set according to actual application needs, which is not limited here.

[0086] Different from the aforementioned embodiment, the first feature is obtained by performing feature extraction on the test status information, and the second feature is obtained by performing feature extraction on the test image, so as to fuse the first feature and the second feature to obtain the fused feature of the test status information and the test image, and then the fused feature is used to perform prediction to obtain the action to be performed by the robot arm at the second moment, which can help improve the prediction accuracy.

[0087] See also Figure 4 , Figure 4 This is a flow chart of another embodiment of the breakthrough point test of the present application. Specifically, it may include the following steps:

[0088] Step S41: acquiring the test state information when the robot arm does not perform the target action as the first state information, and acquiring the test state information after the robot arm completes the target action as the second state information.

[0089] In an implementation scenario, the test status information when the robot arm does not perform the target action may include at least one of the following: joint posture information and ambient light information when the robot arm does not perform the target action; the test status information after the robot arm completes the target action may include at least one of the following: joint posture information and ambient light information after the robot arm completes the target action. The joint posture information and ambient light information can be specifically referred to the relevant description in the aforementioned disclosed embodiment, which will not be repeated here.

[0090] For the convenience of description, in the embodiment of the present disclosure, the target action at the first moment is recorded as a.t , the robot arm did not perform the target action a t The test status information at that time, that is, the first status information is recorded as s t , the robot arm completes the target action a t The subsequent test status information, i.e. the second status information, is recorded as s t+1 In addition, in the embodiment of the present disclosure, the target action a t It can be considered as the action space information of the robot, and the test state information can be considered as the state space information of the robot.

[0091] Step S42: Obtaining the liveness detection result of the liveness attack sample.

[0092] In the disclosed embodiment, the live attack sample may be provided on a robotic arm, and the detection result may be obtained by detecting a test image of the live attack sample, and the test image is obtained by photographing the live attack sample after the robotic arm completes the target action at the first moment.

[0093] For details, please refer to the relevant steps in the aforementioned disclosed embodiments, which will not be repeated here.

[0094] Step S43: using the liveness detection result, obtaining the optimized excitation value of the first neural network.

[0095] In the disclosed embodiment, the optimized incentive value represents a reward for the first neural network. When the live attack sample is detected as non-live, the optimized incentive value is a non-positive value. When the live attack sample is detected as live, the optimized incentive value is a positive value. The live detection result includes a detection probability value, and the optimized incentive value is positively correlated with the detection probability value.

[0096] In an implementation scenario, the detection probability value is used to indicate the probability of being detected as a live body. Specifically, a larger detection probability value indicates a smaller probability, and vice versa. A smaller detection probability value indicates a larger probability. Specifically, a probability threshold can be pre-set to determine whether a live attack sample is detected as a live body or a non-live body based on the size relationship between the detection probability value and the probability threshold. For details, please refer to the relevant description in the aforementioned disclosed embodiment, which will not be repeated here. In the disclosed embodiment, for the convenience of description, the detection probability value is recorded as score, and the target action a at the first moment is completed. t Afterwards, the obtained optimized incentive value can be recorded as r t+1 .

[0097] In a specific implementation scenario, the detection probability value is within a second numerical range, and the second numerical range can be set according to actual application needs, for example, it can be set to 0 to 1, which is not limited here.

[0098] In another specific implementation scenario, the detection of a live attack sample as a live body can be determined based on the detection probability value, and the optimized incentive value can be determined as the product of the detection probability value and the first value and the sum of the second value. Specifically, the first value and the second value are both positive numbers, and the first value is not greater than the second value. In order to avoid reducing the impact of the detection probability value on the optimized incentive value due to the first value being too small, the first value and the second value can be the same or similar in magnitude. For example, the first value and the second value can both be set to 100, or the first value can be set to 990, the second value can be set to 1000, and so on, which are not limited here. In order to simplify the calculation method of the optimized incentive value, the first value and the second value can both be set to 1, and then the optimized incentive value r can be obtained using the following formula t+1 :

[0099] r t+1 =score+1……(1)

[0100] In another specific implementation scenario, the detection of a live attack sample as a non-live object can be determined based on the detection probability value, and the optimized incentive value is determined as the difference between the product of the detection probability and the first value and the second value. The setting method of the first value and the second value can refer to the above description, which will not be repeated here. When the first value and the second value are both set to 1, the optimized incentive value r can be obtained by the following formula: t+1 :

[0101] r t+1 =score-1……(2)

[0102] In the above method, when the detection probability value is greater than the probability threshold, the larger the detection probability value, the less likely the live attack sample is to be detected as live. At this time, the optimized incentive value r t+1 is negative, and the smaller its absolute value is, the smaller the reward given to the first neural network is; or, when the detection probability value is greater than the probability threshold, the smaller the detection probability value is, the more likely the live attack sample is to be detected as live, and the incentive value r is optimized at this time. t+1 is negative, and the larger its absolute value is, the greater the reward given to the first neural network. Therefore, by setting the optimized incentive value of the first neural network, since the first neural network tends to obtain higher rewards during the action prediction process, when the detection probability value becomes larger (that is, the reward becomes smaller), the first neural network can adjust the action prediction strategy and predict the action in the direction of increasing the reward, thereby reducing the detection probability value, that is, predicting the action in the direction of the living attack sample being detected as a living body. Therefore, setting the optimized incentive value in the above manner can stimulate the first neural network, reduce the probability of falling into a non-optimal solution by performing actions back and forth at an unbroken position, and further improve the speed of the breakthrough point test.

[0103] Step S44: using the first state information, the target action, the optimized incentive value and the second state information as training samples to train the first neural network.

[0104] Please refer to Figure 5 , Figure 5 s is a schematic diagram of a state of an embodiment of the present application's breakthrough point test. In the present disclosure embodiment, the first state information s t , target action a t , optimize the incentive value r t+1 and the second state information s t+1 As training samples, the first neural network is trained. In the embodiment of the present disclosure, as mentioned above, the test state information can be considered as the state space information of the robot arm, and the target action a t It can be considered as the action space information of the robot arm, and the optimized incentive value r t+1 It can be considered as the target action a t Therefore, the state space information, action space information and behavior reward can be used as a set of encoded state sequences to jointly train the first neural network.

[0105] In one implementation scenario, the first state information s t It may further include: not executing the target action a t When the live attack sample is tested for liveness detection, the second state information s t+1 It may further include: completing the target action a t Afterwards, the test images taken for liveness detection of the liveness attack samples are taken.

[0106] In another implementation scenario, the first state information, the target action, the optimized incentive value, and the second state information can be used as training samples corresponding to the first moment and saved. After saving the training samples, the first neural network is not trained first, but step S45 is performed first to obtain the action to be executed at the second moment. t+1 , and repeat multiple rounds of steps S41 to S44 to obtain training samples corresponding to multiple moments, until the number of stored training samples exceeds a preset number (e.g., 50, 100, 200, etc.), and then divide the stored training samples into multiple batches, and use the training samples of each batch to train the first neural network and adjust the network parameters of the first neural network.

[0107] In another specific implementation scenario, for ease of description, the first neural network can be denoted as Q, the network parameter of the first neural network can be denoted as θ, and the first state information s can be denoted as t Input the target action a predicted by the first neural network tThe corresponding predicted probability value Q(s t ,a t ; θ), as training samples (s t ,a t ,r t+1 ,s t+1 ) and use the optimized incentive value r t+1 and the second state information s t+1 , determine the training samples (s t ,a t ,r t+1 ,s t+1 )’s target value y t , please refer to the following formula for details:

[0108]

[0109] When the target action a is completed t The second state information s t+1 When it is the breakthrough point information, the target value y t The optimized incentive value r t+1 , when the target action a is completed t The second state information s t+1 When it is not the breakthrough point information, the target value y t The optimized incentive value r t+1 +γmax a′ Q(s t+1 ,a′;θ), where γ represents the discount factor, max a′ Q(s t+1 ,a′;θ) represents the second state information s t+1 The maximum predicted probability value predicted by the first neural network is input, and the candidate action corresponding to the maximum predicted probability value is a′. The candidate action can be specifically represented by numbers. For example, the moving distance and rotation angle of the robot joint can be represented by numbers, so that different candidate actions can be represented by different combinations of numbers. In addition, the larger the discount coefficient γ, the more the breakthrough point test focuses on future predictions. On the contrary, the smaller the discount coefficient γ, the more the breakthrough point test focuses on current incentives. It can be set according to actual application needs and is not limited here. After obtaining the predicted value Q(s t ,a t ;θ) and target value y t After that, we can use the objective function (y t -Q(s t ,a t ;θ)) 2 Perform gradient descent optimization on the first neural network.

[0110] In another specific implementation scenario, in the above training process, using the same first neural network to obtain the target value and predicted value of the training sample may enhance the correlation between the target value and the predicted value, resulting in a greater possibility of oscillation and divergence of the first neural network. In order to improve the accuracy of model training, a target model can be initialized. Target Model The network structure is the same as the first neural network, and the target model is initialized In order to confuse the network parameters θ of the first neural network Q with the network parameters θ of the first neural network Q, the network parameters of the first neural network Q are recorded as θ - Therefore, the training sample (s t ,a t ,r t+1 ,s t+1 ) can be expressed as Training samples (s t ,a t ,r t+1 ,s t+1 )’s target value y t , please refer to the following formula for details:

[0111]

[0112] Getting the predicted value and the target value y t Afterwards, we can use the objective function Perform gradient descent optimization on the first neural network and use the target model every preset number of rounds Update the network parameters θ of the first neural network Q - .

[0113] Step S45: Use the first neural network to perform predictive processing on the test state information and the test image to obtain the action to be performed by the robotic arm at the second moment.

[0114] For details, please refer to the relevant steps in the aforementioned disclosed embodiments, which will not be repeated here.

[0115] The execution order of the above steps S44 and S45 can be set according to the actual application needs. For example, in order to increase the first neural network's attention to the liveness detection result generated by the robot arm at the first moment just passed, step S44 can be executed first, and then step S45. Or, in order to increase the prediction speed of the action to be executed at the second moment, step S45 can be executed first, and then step S44. There is no limitation here. In addition, in order to improve the efficiency of the breakthrough point test, if there are sufficient computing resources, the above steps S44 and S45 can also be executed simultaneously.

[0116] Please continue to refer to Figure 5 In an implementation scenario, after obtaining the action to be executed by the robot arm at the second moment, the second moment can be used as the new first moment, and the action to be executed at the second moment can be used as the target action of the new first moment, and the above step S41 and subsequent steps can be re-executed, so that the process of predicting the action to be executed at the second moment and training the first neural network using training samples can be repeated, so as to continuously optimize the strategy of the robot arm to find the breakthrough point.

[0117] Different from the above-mentioned embodiment, the test state information when the robot arm does not perform the target action is obtained as the first state information, and the test state information after the robot arm completes the target action is used as the second state information, and the liveness detection result is used to obtain the optimized incentive value of the first neural network, and when the liveness attack sample is detected as non-live, the optimized incentive value is a non-positive value, and when the liveness attack sample is detected as live, the optimized incentive value is a positive value, and the liveness detection result includes a detection probability value, and the detection probability value is positively correlated with the optimized incentive value, so that the first state information, the target action, the optimized incentive value and the second state information are used as training samples to train the first neural network, which can be beneficial to the first neural network in the process of predicting the action to be executed at the second moment, and because when the liveness attack sample is detected as non-live, the optimized incentive value is a non-positive value, and when the liveness attack sample is detected as live, the optimized incentive value is a positive value, it can be beneficial to make the first neural network learn the strategy of finding the breakthrough point more quickly according to the positive and negative and size of the optimized incentive value, which is beneficial to improve the speed of the breakthrough point test.

[0118] See also Figure 6 , Figure 6 : is a schematic diagram of a framework of an embodiment of a breach point test device 60 of the present application. The breach point test device 60 includes a state information acquisition module 61, a detection result acquisition module 62 and a prediction module 63. The state information acquisition module 61 is used to acquire the test state information of the robot arm after completing the target action at the first moment; wherein the test state information includes the joint posture information of the robot arm; the detection result acquisition module 62 is used to acquire the liveness detection result of the liveness attack sample, wherein the liveness attack sample is set on the robot arm; the prediction module 63 is used to obtain the breach point information of the liveness detection and the action to be performed by the robot arm at the second moment based on the liveness detection result and the test state information.

[0119] The above scheme obtains the test state information of the robot arm after completing the target action at the first moment, and the test state information includes the joint posture information of the robot arm, and obtains the liveness detection result of the liveness attack sample, and the liveness attack sample is set on the robot arm, so as to obtain the breakthrough point information of the liveness detection and the action to be executed by the robot arm at the second moment based on the liveness detection result and the test state information, so it is possible to automatically find the breakthrough point information by relying on the robot arm without manpower, so it is possible to improve the efficiency of the breakthrough point test. In addition, because it is possible to predict the action to be executed at the first moment, the next moment and the second moment based on the liveness detection result and the test state information of completing the target action at the first moment, it is possible to avoid time-consuming traversal of all pre-designed actions during the breakthrough point test, and actively predict the action to be executed at the next moment based on the current liveness detection result and the test state information, so that the movement strategy of the robot arm is more targeted, which can be conducive to improving the speed of the breakthrough point test.

[0120] In some disclosed embodiments, the breakthrough point testing device 60 also includes a target action acquisition module, which is used to obtain the target action of the robotic arm at the first moment based on the target liveness detection result and target test status information of the robotic arm at the previous moment before the first moment. The breakthrough point testing device 60 also includes an action control module, which is used to control the robotic arm to complete the target action at the first moment.

[0121] Different from the aforementioned embodiment, similar to the action to be performed by the robotic arm at the second moment, the target action completed by the robotic arm at the first moment is obtained by using the target liveness detection result and target test status information at the previous moment before the first moment, which can be beneficial for continuously looping execution actions and prediction actions during the entire breakthrough point test process, thereby improving the efficiency and robustness of the breakthrough point test.

[0122] In some disclosed embodiments, the detection result is obtained by detecting a test image of a live attack sample, and the prediction module 63 includes a first acquisition submodule, which is used to obtain the action to be performed by the robotic arm at the second moment according to the test status information and the test image when the liveness detection result of the liveness attack sample is non-liveness; wherein the test image is obtained by shooting the liveness attack sample after the robotic arm completes the action to be performed at the first moment.

[0123] Different from the aforementioned embodiment, the detection result is set to be obtained by detecting a test image of a live attack sample, so that when the liveness detection result of the liveness attack sample is non-liveness, the action to be performed by the robot arm at the second moment is obtained according to the test state information and the test image, and the test image is obtained by photographing the liveness attack sample after the robot arm completes the action to be performed at the first moment, and then the action to be performed by the robot arm at the second moment can be obtained by combining the test state information and the real-world test image, thereby reducing the error rate caused by predicting the action to be performed due to relying on a single information, which can be beneficial to improving the prediction accuracy.

[0124] In some disclosed embodiments, the first acquisition submodule includes a first feature extraction unit, which is used to perform feature extraction on the test status information to obtain a first feature, and the first acquisition submodule includes a second feature extraction unit, which is used to perform feature extraction on the test image to obtain a second feature; the first acquisition submodule includes a feature fusion unit, which is used to fuse the first feature and the second feature to obtain a fusion feature of the test status information and the test image; the first acquisition submodule includes a prediction unit, which is used to obtain the action to be performed by the robot arm at the second moment based on the fusion feature.

[0125] Different from the aforementioned embodiment, the first feature is obtained by performing feature extraction on the test status information, and the second feature is obtained by performing feature extraction on the test image, thereby fusing the first feature and the second feature to obtain the fused feature of the test status information and the test image, and then based on the fused feature, the action to be performed by the robot arm at the second moment is obtained. Therefore, the feature of the test status information and the feature of the real-world test image can be combined to obtain the action to be performed, which can help improve the accuracy of the action to be performed.

[0126] In some disclosed embodiments, the first feature extraction unit is specifically used to perform feature extraction on the test state information using the first extraction subnetwork of the first neural network to obtain the first feature; the second feature extraction unit is specifically used to perform feature extraction on the test image using the second extraction subnetwork of the first neural network to obtain the second feature; the prediction unit is specifically used to predict the fusion feature using the prediction subnetwork of the first neural network to obtain the action to be performed by the robotic arm at the second moment.

[0127] Different from the aforementioned embodiment, the first extraction subnetwork of the first neural network is used to perform feature extraction on the test state information to obtain the first feature, and the second extraction subnetwork of the first neural network is used to perform feature extraction on the test image to obtain the second feature, which can help improve the efficiency of feature extraction, and the prediction subnetwork of the first neural network is used to predict the fusion feature to obtain the action to be performed by the robot arm at the second moment. Therefore, the first neural network can be used to realize feature extraction and action prediction, which can help improve the efficiency of the breakthrough point test.

[0128] In some disclosed embodiments, the prediction unit includes a probability acquisition subunit, which is used to process the fused features using a prediction subnetwork to obtain predicted probability values ​​of multiple candidate actions; the prediction unit includes an action determination subunit, which is used to use the candidate action corresponding to the largest predicted probability value as the action to be performed by the robotic arm at the second moment.

[0129] Different from the aforementioned embodiment, by using the prediction sub-network to process the fusion features, the predicted probability values ​​of multiple candidate actions are obtained, and the candidate action corresponding to the largest predicted probability value is used as the action to be executed by the robotic arm at the second moment, which can help improve the success rate of the breakthrough point test.

[0130] In some disclosed embodiments, the test state information also includes ambient lighting information, and the first feature extraction unit is specifically used to use the first extraction sub-network to perform matrix multiplication on the joint posture information and the ambient lighting information to obtain the first feature.

[0131] Different from the aforementioned embodiment, when the test status information also includes ambient lighting information, the first extraction subnetwork is used to perform matrix multiplication on the joint posture information and the ambient lighting information to obtain the first feature, which can be helpful in fusing the ambient lighting information and the joint posture information, thereby helping to improve the feature richness of the first feature, and further can help to improve the accuracy of subsequent action prediction and improve the success rate of breakthrough point testing.

[0132] In some disclosed embodiments, the action to be performed by the robot arm at the second moment is predicted by the first neural network, and the state information acquisition module 61 is also used to obtain test state information when the robot arm has not performed the action to be performed, as the first state information, and the test state information after the robot arm completes the action to be performed is used as the second state information. The breakthrough point testing device 60 also includes an optimized incentive acquisition module, which is used to use the liveness detection result to obtain the optimized incentive value of the first neural network; wherein, when the liveness attack sample is detected as non-liveness, the optimized incentive value is a non-positive value, and when the liveness attack sample is detected as liveness, the optimized incentive value is a positive value, and the liveness detection result includes a detection probability value, and the optimized incentive value is positively correlated with the detection probability value; the breakthrough point testing device 60 also includes a model training module, which is used to use the first state information, the action to be performed, the optimized incentive value and the second state information as training samples to train the first neural network.

[0133] Different from the above-mentioned embodiment, the test state information when the robot arm does not perform the action to be performed is obtained as the first state information, and the test state information after the robot arm completes the action to be performed is used as the second state information, and the liveness detection result is used to obtain the optimized incentive value of the first neural network, and when the liveness attack sample is detected as non-live, the optimized incentive value is a non-positive value, and when the liveness attack sample is detected as live, the optimized incentive value is a positive value, and the liveness detection result includes a detection probability value, and the optimized incentive value and the detection probability value are positively correlated, so that the first state information, the action to be performed, the optimized incentive value and the second state information are used as training samples to train the first neural network, which can be beneficial to the first neural network in the process of predicting the action to be performed at the second moment, and because when the liveness attack sample is detected as non-live, the optimized incentive value is a non-positive value, and when the liveness attack sample is detected as live, the optimized incentive value is a positive value, it can be beneficial to make the first neural network learn the strategy of finding the breakthrough point more quickly according to the positive and negative and size of the optimized incentive value, which is beneficial to improve the speed of the breakthrough point test.

[0134] In some disclosed embodiments, the optimized incentive acquisition module includes a first acquisition submodule, which is used to determine that the live attack sample is detected as live based on the detection probability value, and determine the optimized incentive value as the product of the detection probability value and the first numerical value and the sum of the second numerical value; and / or, the optimized incentive acquisition module includes a second acquisition submodule, which is used to determine that the live attack sample is detected as non-live based on the detection probability, and determine the optimized incentive value as the difference between the product of the detection probability value and the first numerical value and the second numerical value; wherein the first numerical value and the second numerical value are both positive numbers, and the first numerical value is not greater than the second numerical value.

[0135] Different from the aforementioned embodiment, when the live attack sample is detected as a live body based on the detection probability value, the optimized incentive value is determined as the product of the detection probability value and the first value and the sum of the second value, and when the live attack sample is detected as a non-live body based on the detection probability, the optimized incentive value is determined as the difference between the product of the detection probability value and the first value and the second value, and the first value and the second value are both positive numbers, and the first value is not greater than the second value. Therefore, when the live attack sample is detected as a non-live body based on the detection probability, the optimized incentive value is calculated as a non-positive value, thereby stimulating the first neural network, which can reduce the probability of falling into a non-optimal solution by performing actions back and forth at an unbroken position, and further improve the speed of the breakthrough point test.

[0136] In some disclosed embodiments, the prediction module 63 includes a second acquisition submodule, which is used to reset the joint posture of the robotic arm to a preset initial value and acquire the preset initial action as the action to be performed by the robotic arm at the second moment when the liveness detection result of the liveness attack sample is liveness.

[0137] Different from the aforementioned embodiment, when the liveness detection result of the liveness attack sample is liveness, by resetting the joint posture of the robotic arm to a preset initial value and obtaining the preset initial action as the action to be executed by the robotic arm at the second moment, the robotic arm can start a new round of breakthrough point testing, which can help increase the amount of breakthrough point information obtained from the test.

[0138] In some disclosed embodiments, the prediction module 63 is further specifically configured to determine the test status information as the breach point information when the liveness detection result of the liveness attack sample is liveness.

[0139] Different from the above-mentioned embodiment, when the live attack sample is detected as live, the test status information is determined as the breach point information, which can be helpful for automatically saving the breach point information in a timely manner during the breach point test process, thereby being helpful for optimizing and improving the liveness detection later.

[0140] In some disclosed embodiments, the liveness detection result is obtained by using the second neural network detection, and the breakthrough point testing device 60 also includes a network optimization module for optimizing the second neural network using the breakthrough point information.

[0141] Different from the aforementioned embodiment, the second neural network used for living body detection is optimized through the breakthrough point information obtained through the breakthrough point test, which can be beneficial to improving the performance of the second neural network.

[0142] See also Figure 7 , Figure 7: is a schematic diagram of a framework of an embodiment of an electronic device 70 of the present application. The electronic device 70 includes a memory 71 and a processor 72 coupled to each other, and the processor 72 is used to execute program instructions stored in the memory 71 to implement the steps of any of the above-mentioned breach point test method embodiments. In a specific implementation scenario, the electronic device 70 may include but is not limited to: a microcomputer, a laptop computer, a tablet computer and other devices, which are not limited here.

[0143] Specifically, the processor 72 is used to control itself and the memory 71 to implement the steps of any of the above-mentioned breach point test method embodiments. The processor 72 may also be referred to as a CPU (Central Processing Unit). The processor 72 may be an integrated circuit chip having signal processing capabilities. The processor 72 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. In addition, the processor 72 may be implemented by an integrated circuit chip.

[0144] The above solution can improve the efficiency of the breakthrough point test.

[0145] See also Figure 8 , Figure 8 The schematic diagram of the framework of an embodiment of a computer-readable storage medium 80 of the present application. The computer-readable storage medium 80 stores program instructions 801 that can be executed by a processor, and the program instructions 801 are used to implement the steps of any of the above-mentioned breach point testing method embodiments.

[0146] The above solution can improve the efficiency of the breakthrough point test.

[0147] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0148] The above description of various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, they will not be repeated herein.

[0149] In the several embodiments provided in the present application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation described above is only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation, such as units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0150] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0151] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0152] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) to perform all or part of the steps of each implementation method of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

Claims

1. A method for testing a breach point, characterized in that: include: Acquire test state information of the robot arm completing the target action at the first moment; wherein the test state information includes joint posture information of the robot arm; Obtaining a liveness detection result of a liveness attack sample; wherein the liveness attack sample is provided on the robotic arm; When the liveness detection result is liveness, the test state information is determined as the breakthrough point information of the liveness detection; when the liveness detection result is non-liveness, the action to be performed by the robot arm at a second moment is predicted based at least on the test state information; wherein the second moment is the next moment of the first moment; The liveness detection result is obtained by detecting a test image of the liveness attack sample. When the liveness detection result of the liveness attack sample is non-liveness, predicting the action to be performed by the robot arm at the second moment based at least on the test state information includes: Performing feature extraction on the test state information to obtain a first feature, and performing feature extraction on the test image to obtain a second feature; The first feature and the second feature are fused to obtain a fused feature of the test state information and the test image; According to the fusion feature, the action to be performed by the robotic arm at the second moment is obtained; wherein the test image is obtained by photographing the living attack sample after the robotic arm completes the target action.

2. The method according to claim 1, characterized in that Before obtaining the test state information of the robot arm completing the target action at the first moment, the method includes: Based on the target liveness detection result and target test state information of the robotic arm at the last moment before the first moment, obtaining the target action of the robotic arm at the first moment; The robotic arm is controlled to complete the target action at the first moment.

3. The method according to claim 1, characterized in that The extracting features of the test status information to obtain a first feature includes: Using a first extraction sub-network of a first neural network to extract features from the test state information, to obtain the first feature; And / or, extracting features from the test image to obtain a second feature comprises: Using the second extraction subnetwork of the first neural network to extract features from the test image, to obtain the second features; And / or, obtaining the action to be performed by the robot arm at the second moment according to the fusion feature includes: The prediction subnetwork of the first neural network is used to predict the fusion feature to obtain the action to be performed by the robotic arm at the second moment.

4. The method according to claim 3, characterized in that The using the prediction sub-network of the first neural network to predict the fusion feature to obtain the action to be performed by the robot arm at the second moment includes: Processing the fused features using the prediction subnetwork to obtain predicted probability values ​​of multiple candidate actions; The candidate action corresponding to the maximum predicted probability value among the predicted probability values ​​is used as the action to be performed by the robotic arm at the second moment.

5. The method according to claim 3 or 4, characterized in that: The test state information also includes ambient light information, and the first extraction sub-network of the first neural network is used to extract features from the test state information to obtain the first feature, including: The first extraction subnetwork is used to perform matrix multiplication on the joint posture information and the ambient light information to obtain the first feature.

6. The method according to claim 1, characterized in that The action to be performed by the robotic arm at the second moment is predicted by the first neural network, and the method further includes: Acquire test state information when the robot arm does not perform the target action as first state information, and acquire test state information after the robot arm completes the target action as second state information; and Using the liveness detection result, obtaining an optimized excitation value of the first neural network; wherein, when the liveness attack sample is detected as non-liveness, the optimized excitation value is a non-positive value, and when the liveness attack sample is detected as liveness, the optimized excitation value is a positive value, and the liveness detection result includes a detection probability value, and the optimized excitation value is positively correlated with the detection probability value; The first state information, the target action, the optimized incentive value and the second state information are used as training samples to train the first neural network.

7. The method according to claim 6, characterized in that The step of obtaining the optimized excitation value of the first neural network by using the liveness detection result includes: Determine that the live attack sample is detected as live based on the detection probability value, and determine the optimized incentive value as the product of the detection probability value and the first value and the sum of the second value; and / or, Determine, based on the detection probability, that the live attack sample is detected as a non-live body, and determine the optimized incentive value as a difference between a product of the detection probability value and a first value and a second value; The first value and the second value are both positive numbers, and the first value is not greater than the second value.

8. The method according to claim 1, characterized in that When the liveness detection result of the liveness attack sample is liveness, the method further includes: The joint posture of the robotic arm is reset to a preset initial value, and a preset initial action is obtained as the action to be performed by the robotic arm at the second moment.

9. The method according to claim 1, characterized in that: The liveness detection result is obtained by using a second neural network detection; the method further includes: The second neural network is optimized using the breakthrough point information.

10. A breach point testing device, characterized in that: include: A state information acquisition module, used to acquire test state information of the robot arm completing a target action at a first moment; wherein the test state information includes joint posture information of the robot arm; A detection result acquisition module, used to obtain a liveness detection result of a liveness attack sample; wherein the liveness attack sample is arranged on the mechanical arm; A prediction module, configured to, when the liveness detection result is liveness, determine the test state information as the breakthrough point information of the liveness detection, and when the liveness detection result is non-liveness, predict the action to be performed by the robot arm at a second moment based at least on the test state information; wherein the second moment is a moment after the first moment; The liveness detection result is obtained by detecting a test image of the liveness attack sample. When the liveness detection result of the liveness attack sample is non-liveness, predicting the action to be performed by the robot arm at the second moment based at least on the test state information includes: Performing feature extraction on the test state information to obtain a first feature, and performing feature extraction on the test image to obtain a second feature; The first feature and the second feature are fused to obtain a fused feature of the test state information and the test image; According to the fusion feature, the action to be performed by the robotic arm at the second moment is obtained; wherein the test image is obtained by photographing the living attack sample after the robotic arm completes the target action.

11. An electronic device, characterized in that: It comprises a memory and a processor, wherein the processor is used to execute program instructions stored in the memory to implement the breakthrough point testing method according to any one of claims 1 to 9.

12. A computer-readable storage medium having program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, the breach point testing method according to any one of claims 1 to 9 is implemented.

Images