Gait identity recognition hidden disturbance detection method and device

Abstract

The application discloses a hidden disturbance detection method and device for gait identity recognition, and relates to the field of artificial intelligence security and biometric recognition protection. The method is used to solve the problem of poor detection sensitivity caused by the difficulty of amplifying micro hidden disturbances in current defense technology. The method comprises the following steps: obtaining cumulative normal gait data and cumulative abnormal disturbance samples in a target scene according to a set period, performing full-dimensional adaptive optimization to obtain an optimal detection threshold, optimal amplification parameters, an optimal encoder matrix and an optimal decoder matrix; obtaining tampered ciphertext data injected with hidden disturbances, inputting the tampered ciphertext data into the optimal decoder matrix to obtain amplified disturbances; determining the L2 norm of the amplified disturbances, and determining that there are hidden disturbances in the tampered ciphertext data when the L2 norm of the amplified disturbances is greater than the optimal detection threshold.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence security and biometric identification protection, and more specifically to a method and apparatus for detecting concealed disturbances in gait identification. Background Technology

[0002] Gait identification technology, as an important branch of biometric identification, has become a core supporting technology for scenarios such as security monitoring, smart park access control, and human-computer interaction identity verification due to its unique advantages such as non-contact data collection, difficulty in spoofing, and long-distance identification. In recent years, it has made significant breakthroughs in algorithm accuracy optimization and adaptability to complex scenarios, and has gradually realized the application of gait identification technology from laboratory verification to large-scale commercial use.

[0003] However, with the widespread deployment of gait identification systems in key areas, their security vulnerabilities and risks of external disturbances are becoming increasingly prominent. On the one hand, as a biometric feature, gait data is susceptible to tampering and manipulation during its collection and transmission by hidden false data injection. In other words, illegal tamperers can alter gait characteristics by injecting minute, hidden disturbances into normal gait data that are difficult for human vision or conventional detection algorithms to perceive, thereby inducing the system to misjudge or forge identities, seriously threatening system security. On the other hand, the decision-making mechanisms of existing gait identification systems mostly rely on black-box reasoning of deep learning models. The sensitivity of the models to minute, hidden disturbances has not been fully considered, resulting in insufficient intrinsic security defense capabilities.

[0004] From the current state of technological research and application, there are still significant shortcomings in the security protection research of gait identity recognition systems: existing defense methods mostly focus on traditional anomaly detection. These methods can only identify abnormal gait data with large amplitude, and have almost no ability to detect tiny hidden perturbations far below the anomaly detection threshold. They cannot trigger alarms from traditional detectors, and can accurately exploit model vulnerabilities to achieve perturbation. At the same time, some studies have attempted to improve the robustness of the system through model adversarial training, but this method relies on a large number of labeled perturbation samples, and has high training costs and poor generalization. It is difficult to adapt to the differences in gait data distribution in different scenarios, and cannot form a universal defense solution.

[0005] In summary, while the field of gait identity recognition has matured in terms of recognition accuracy and scene adaptability, there is a significant lag in defense technologies against concealed disturbances. Traditional detection methods are ineffective, and the cost of adversarial training is high. There is an urgent need for a new protection technology that can efficiently amplify small concealed disturbances while balancing detection accuracy and deployment cost, in order to cope with the increasingly severe security threats and ensure the reliable application of gait identity recognition systems in critical scenarios. Summary of the Invention

[0006] This invention provides a method and apparatus for detecting concealed perturbations in gait identification, which addresses the problem of poor detection sensitivity in current defense technologies due to the difficulty in amplifying minute concealed perturbations.

[0007] This invention provides a method for detecting occult perturbations in gait identification, comprising:

[0008] According to the set period, the accumulated normal gait data and accumulated abnormal disturbance samples in the target scene are subjected to full-dimensional adaptive optimization to obtain the optimal detection threshold, optimal amplification parameters, optimal encoder matrix and optimal decoder matrix.

[0009] The process involves: acquiring initial gait feature data of a normal user in a target scenario; inputting the initial gait feature data into the optimal encoder matrix to obtain initial ciphertext data; acquiring tampered ciphertext data injected with hidden perturbations; inputting the tampered ciphertext data into the optimal decoder matrix to obtain amplified perturbations; wherein the tampered ciphertext data includes the initial ciphertext data and the hidden perturbations; and the amplified perturbations are obtained after the hidden perturbations are amplified.

[0010] The L2 norm of the amplified perturbation is determined. When the L2 norm of the amplified perturbation is greater than the optimal detection threshold, it is determined that there is a hidden perturbation in the tampered ciphertext data, and the hidden perturbation does not match the perturbation pattern included in the accumulated abnormal perturbation samples.

[0011] This invention provides a concealed perturbation detection device for gait identity recognition, comprising:

[0012] The first acquisition unit is used to perform full-dimensional adaptive optimization on the accumulated normal gait data and accumulated abnormal disturbance samples in the acquired target scene according to a set period, so as to obtain the optimal detection threshold, optimal amplification parameters, optimal encoder matrix and optimal decoder matrix.

[0013] An amplification unit is used to acquire initial gait feature data of a normal user in a target scene, input the initial gait feature data into the optimal encoder matrix to obtain initial ciphertext data; acquire tampered ciphertext data injected with hidden perturbations, input the tampered ciphertext data into the optimal decoder matrix to obtain amplified perturbations; wherein, the tampered ciphertext data includes the initial ciphertext data and the hidden perturbations; the hidden perturbations are amplified to obtain the amplified perturbations;

[0014] The judgment unit is used to determine the L2 norm of the amplified perturbation. When the L2 norm of the amplified perturbation is greater than the optimal detection threshold, it is determined that there is a hidden perturbation in the tampered ciphertext data, and the hidden perturbation does not match the perturbation pattern included in the accumulated abnormal perturbation samples.

[0015] This invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform any of the above-described methods for gait identity recognition of concealed perturbation detection.

[0016] This invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs any of the above-described methods for gait identity recognition to detect hidden perturbations.

[0017] This invention provides a method and apparatus for detecting occult perturbations in gait identification. The method performs full-dimensional optimization on accumulated data at a set period, outputting optimal parameters and matrices that adapt to the noise and perturbation features of the current scene, ensuring detection accuracy from the source and avoiding detection failures caused by changes in data distribution and iterative perturbation patterns in static frameworks. Furthermore, relying on encoder and decoder matrices constructed from orthogonal and diagonal matrices, it achieves directional perturbation amplification of core gait features, transforming minute occult perturbations imperceptible to human vision and conventional detection into quantifiable amplified perturbations while maintaining normal gait. The data is encoded and decoded without loss of quality, without affecting the original authentication function of the system. Furthermore, the detection threshold and amplification parameters are dynamically set based on the actual data of the scene, accurately distinguishing between normal noise and abnormal disturbances, effectively reducing the probability of false detection and false negative detection, and the parameters are adapted to different application scenarios, improving the generalization of the solution. By classifying and updating the accumulated normal gait data and accumulated abnormal disturbance samples through detection results, the latest real data is provided for the next round of optimization, forming a positive closed loop of optimization, detection, data accumulation, and re-optimization, allowing the detection capability to continuously iterate with changes in the scene and the evolution of disturbances, solving the problem of lagging feature library updates in traditional solutions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic flowchart of a concealed perturbation detection method for gait identity recognition provided in an embodiment of the present invention;

[0020] Figure 2 This is a schematic diagram of the concealed disturbance detection device for gait identity recognition provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Figure 1 This is a schematic flowchart of a concealed perturbation detection method for gait identity recognition provided in an embodiment of the present invention; as shown below. Figure 1 As shown, the method includes the following steps:

[0023] Step 101: According to the set period, perform full-dimensional adaptive optimization on the accumulated normal gait data and accumulated abnormal perturbation samples in the acquired target scene to obtain the optimal detection threshold, optimal amplification parameters, optimal encoder matrix and optimal decoder matrix.

[0024] Step 102: Obtain initial gait feature data of a normal user in the target scene, and input the initial gait feature data into the optimal encoder matrix to obtain initial ciphertext data; obtain tampered ciphertext data injected with hidden perturbations, and input the tampered ciphertext data into the optimal decoder matrix to obtain amplified perturbations; wherein, the tampered ciphertext data includes the initial ciphertext data and the hidden perturbations; the hidden perturbations are amplified to obtain the amplified perturbations;

[0025] Step 103: Determine the L2 norm of the amplified perturbation. When the L2 norm of the amplified perturbation is greater than the optimal detection threshold, it is determined that there is a hidden perturbation in the tampered ciphertext data, and the hidden perturbation does not match the perturbation pattern included in the accumulated abnormal perturbation samples.

[0026] The execution subject of the method provided in this embodiment of the invention is the processor included in the gait recognition system.

[0027] In step 101, accumulated normal gait data and accumulated abnormal perturbation samples in the target scene are collected according to a set period, and full-dimensional adaptive optimization is performed to obtain the optimal detection threshold. Optimal amplification parameters Optimal encoder matrix and the optimal decoder matrix .

[0028] Specifically, the set period here refers to the pre-set optimization time interval for the gait recognition system. This period setting breaks the limitations of the traditional static detection framework, avoiding detection failures caused by dynamic changes in gait data distribution and continuous iteration of hidden disturbances. Furthermore, the optimization time interval does not mean immediate optimization after a single detection. The optimization time of the set period takes into account both the sufficiency of data accumulation and the resource consumption of the gait recognition system, adapting to the real-time and stability requirements of security scenarios. In practical applications, the set period for security scenarios can be 7 days, and the set period for smart parks can be 15 days. If the set period is 7 days, the gait recognition system will automatically trigger this full-dimensional adaptive optimization after the end of each 7-day detection period, rather than optimizing every day / every detection.

[0029] Furthermore, the accumulated normal gait data refers to the sum of all normal gait data that have been detected and determined to be free of hidden disturbances in the target scenario, from system startup to the end of the current cycle. It includes initial historical normal data and newly added normal gait data in each round, which come from different user groups, different collection environments and different device states, and reflect the true noise level of the current scenario.

[0030] The cumulative abnormal disturbance sample also refers to the sum of all abnormal disturbance samples that have been detected and determined to have "hidden disturbances" in the target scenario from system startup to the end of the current cycle. It includes the initial abnormal disturbance and the hidden disturbances added in each round. The cumulative abnormal disturbance sample covers all the disturbance patterns that have been discovered and reflects the current tampering disturbance characteristics.

[0031] Furthermore, full-dimensional adaptive optimization refers to simultaneously optimizing the detection threshold, amplification parameters, encoder matrix, and decoder matrix, and the optimization process is completed automatically based entirely on the input accumulated normal gait data and accumulated abnormal perturbation samples.

[0032] Specifically, by calculating the L2 norm of the inherent noise of the accumulated normal gait data, the maximum value of the noise L2 norm is taken. Then, the data obtained is adjusted upwards by 10%-20% to determine the optimal detection threshold. This upward adjustment is based on the actual noise of the target scene, thus accurately distinguishing between normal noise and amplified disturbances. This avoids false detections caused by amplified disturbances exceeding the fixed threshold, or missed detections caused by amplified disturbances not exceeding the fixed threshold, when the optimal detection threshold is a fixed threshold. The 10%-20% upward adjustment range provides adaptability for different scenarios, improving the generalization ability of the solution.

[0033] The optimal amplification parameter is determined based on all perturbation patterns included in the accumulated abnormal perturbation samples and the optimal detection threshold. The weaker the hidden perturbation, the smaller the value of the optimal amplification parameter, and the optimal amplification parameter is greater than zero and less than 1. In this embodiment of the invention, the reciprocal of the optimal amplification parameter is the upper limit of the perturbation amplification factor. Therefore, the smaller the value of the optimal amplification parameter, the larger the corresponding amplification factor, ensuring that all amplified hidden perturbations can be identified by the threshold. By determining the value of the optimal amplification parameter through the above dual constraints, it is ensured that all known hidden perturbations can be detected after amplification, while avoiding excessive amplification of normal noise. The constraint that the optimal amplification parameter is greater than zero and less than 1 ensures directional amplification of perturbations, and the amplification factor can be flexibly adjusted according to the weakness of the perturbation.

[0034] The feature matrix of the accumulated normal gait data is decomposed by QR to obtain an orthogonal matrix that meets the set conditions. The orthogonal matrix is ​​generated by the QR decomposition of the feature matrix of the accumulated normal gait data, which fits the feature dimension distribution of the accumulated normal gait data, ensuring the reversibility and stability of the encoding / decoding process. Moreover, the orthogonal transformation does not change the L2 norm of the data, thus avoiding data distortion.

[0035] Furthermore, the first diagonal element of the diagonal matrix is ​​the optimal amplification parameter, and the remaining diagonal elements are 1, which realizes the directional amplification of the core gait feature dimension without affecting the normal data restoration of other dimensions, thus solving the problem of normal data distortion caused by full-dimensional amplification.

[0036] In this embodiment of the invention, the optimal encoder matrix is ​​determined by the product of an orthogonal matrix, a diagonal matrix, and the transpose of the orthogonal matrix. The inverse of the optimal encoder matrix is ​​the optimal decoder matrix, which can ensure that the initial ciphertext data without hidden disturbances can be 100% restored to the initial gait feature data after passing through the optimal decoder matrix, without affecting the normal authentication function of the gait recognition system.

[0037] Specifically, the optimal encoder matrix, optimal decoder matrix, orthogonal matrix, and diagonal matrix are shown below:

[0038] (1-1)

[0039] (1-2)

[0040] (1-3)

[0041] (1-4)

[0042] in, Represents the optimal encoder matrix. Represents an orthogonal matrix. Represents a diagonal matrix. This represents the transpose of an orthogonal matrix. Represents the optimal decoder matrix. Represents a diagonal matrix The inverse matrix, Represents the identity matrix. This represents the optimal amplification parameter.

[0043] The optimal detection threshold obtained by the above process Optimal amplification parameters Optimal encoder matrix and the optimal decoder matrix As the optimization result within the set period of the gait recognition system, it serves as the initial execution condition and participates in the detection of hidden disturbances within that period. It directly takes the optimal result adapted to the current scenario as the starting point, avoiding the problems of rough parameters and poor adaptability of traditional initial frameworks, and ensuring detection accuracy from the source. At the same time, it provides a benchmark optimized version for subsequent iterative optimization.

[0044] In step 102, the initial gait feature data of a normal user in the target scenario is obtained, and the initial gait feature data is input into the optimal encoder matrix to obtain the initial ciphertext data.

[0045] Specifically, this refers to data that has not been altered by the illegal party and has not been injected with any hidden disturbances. OK The initial gait feature data is fed into the optimal encoder matrix obtained in step 101 as standard input data. Based on the linear transformation function of the optimal encoder matrix, a directional format conversion is performed on the initial gait feature data, converting the intuitive gait feature data into non-intuitive ciphertext data. This blocks the path for illegal tampering parties to directly tamper with the original gait data, increasing the difficulty of perturbation. Through this method, initial ciphertext data with the same dimension as the initial gait feature data is obtained, as shown below:

[0046] (2-1)

[0047] in, Represents the optimal encoder matrix. This represents the initial gait feature data. , This represents the initial ciphertext data. .

[0048] In this embodiment of the invention, the intuitive initial gait feature data is converted into non-intuitive initial ciphertext data through the optimal encoder matrix. On the one hand, this avoids the vulnerability of illegal tamperers directly and maliciously altering the original image without leaving obvious traces, allowing illegal tamperers to inject perturbations into the ciphertext data and preventing them from directly manipulating the original pixel information. On the other hand, this reversible transformation characteristic ensures that the perturbation-free initial ciphertext data can be 100% completely restored to the initial gait feature data after inverse transformation by the optimal decoder matrix, achieving lossless encoding and decoding.

[0049] Furthermore, the tampered ciphertext data injected with a hidden perturbation is obtained, wherein the tampered ciphertext data includes the initial ciphertext data and the hidden perturbation.

[0050] Specifically, the gait recognition system receives tampered ciphertext data transmitted externally or input locally. This type of data is tampered ciphertext data formed by the unauthorized party maliciously injecting hidden perturbations into the initial ciphertext data. In essence, it is a direct superposition and combination of the initial ciphertext data and the hidden perturbations. The data dimension of the tampered ciphertext data is the same as that of the initial ciphertext data. The tampered ciphertext data is shown below:

[0051] (2-2)

[0052] in, This indicates that the encrypted data has been tampered with. This indicates a concealed disturbance. This represents the initial ciphertext data.

[0053] It should be noted that the dimensions of the aforementioned concealed perturbation are consistent with the dimensions of the initial gait feature data. This ensures that the dimensions of the concealed perturbation, amplified perturbation, and initial gait feature data are consistent, guaranteeing the effectiveness of matrix operations. Furthermore, the perturbation is only injected into the dimensions of the initial ciphertext data, which aligns with the actual perturbation method of the illegal tamperer.

[0054] In this embodiment of the invention, since the distribution of gait data changes dynamically and the hidden perturbations are continuously iterated over the period, traditional static frameworks cannot effectively detect and defend against them for a long time.

[0055] Furthermore, the acquired tampered ciphertext data to be detected is input into the optimal decoder matrix obtained in step 101. Using the linear inverse transformation function of the inverse matrix of the optimal decoder matrix, the tampered ciphertext data is decoded to obtain the corrupted plaintext data represented by the amplified perturbation and initial gait feature data. The corrupted plaintext data represented by the optimal decoder matrix and the tampered ciphertext data is shown below:

[0056] (2-3)

[0057] In this embodiment of the invention, the formula for tampering with encrypted data is: The formula for the initial ciphertext data is: Substituting the two formulas above into formula (2-3), we can obtain the corrupted plaintext data based on the optimal decoder matrix, optimal encoder matrix, initial gait feature data, and hidden perturbation representation, as shown below:

[0058] (2-4)

[0059] In this embodiment of the invention, the property of inverse matrix is ​​utilized. Formula (2-4) can be transformed into:

[0060] (2-5)

[0061] Furthermore, let the amplified perturbation be:

[0062] (2-6)

[0063] This yields the corrupted plaintext data based on the amplified perturbation and initial gait feature data, as shown below:

[0064] (2-7)

[0065] Furthermore, the amplified perturbation obtained through formula (2-7) has an amplification factor that satisfies the following formula:

[0066] (2-8)

[0067] In this embodiment of the invention, the constraint of the amplification factor mathematically ensures that the hidden disturbance is effectively amplified, thereby achieving precise and targeted amplification of minute hidden disturbances. Moreover, the amplification factor is controllable, maximizing the matching degree between the amplification effect and the disturbance dimension, avoiding system misjudgment due to excessive amplification caused by matrices not being properly matched, and providing clear quantitative basis for disturbance detection.

[0068] This method achieves directional amplification of hidden perturbations through the inverse transformation of the decoder matrix, transforming minute hidden perturbations imperceptible to human vision / conventional detection into quantifiable and detectable significantly amplified perturbations, thus solving the problems of difficult amplification and low sensitivity of traditional detection perturbations.

[0069] In step 104, the L2 norm of the amplified perturbation is determined, and the L2 norm of the amplified perturbation is compared with the optimal detection threshold determined in step 101.

[0070] When the L2 norm of the amplified perturbation is greater than the optimal detection threshold, it is determined that there is a hidden perturbation in the tampered encrypted data. Then, the newly analyzed hidden perturbation is added to the cumulative abnormal perturbation sample in step 101 to enrich the perturbation pattern library and obtain updated abnormal perturbation samples. This allows the amplification parameters and matrix of the next round of optimization to be adapted to the new perturbation pattern, solving the problem that the feature library update of the traditional scheme lags behind the perturbation evolution.

[0071] When the L2 norm of the amplified perturbation is less than or equal to the optimal detection threshold, it is determined that there is no hidden perturbation in the tampered ciphertext data. The initial gait feature data corresponding to the initial ciphertext data is added to the accumulated normal gait data to obtain updated normal gait data. This allows the noise statistics to continuously fit the dynamic changes of the scene, and the subsequent optimized detection threshold is more accurate.

[0072] In this embodiment of the invention, the finally obtained hidden perturbation is added to the accumulated abnormal perturbation sample to obtain the updated abnormal perturbation sample, or the finally obtained initial gait feature data is added to the accumulated normal gait data to obtain the updated normal gait data. This provides real and up-to-date input data for the next round of full-dimensional adaptive optimization, forming a positive closed loop of detection feedback optimization, which optimizes and improves detection, and ensures the long-term effectiveness of the hidden perturbation detection method for gait identity recognition provided by this embodiment of the invention.

[0073] To more clearly illustrate this invention, the method will be described in detail using third-order gait feature data in a security scenario as the detection object.

[0074] This example selects urban public security monitoring as the core application scenario. This scenario includes sub-applications such as road surveillance checkpoints, community perimeter security, subway / train station passenger flow security checks, and long-distance gait-based turnstiles in industrial parks. These scenarios generally rely on three-dimensional gait features—stride length, stride frequency, and center of gravity height—to complete contactless pedestrian identity verification. Malicious attackers can tamper with gait ciphertext by injecting minute, imperceptible perturbations into the transmission link, thereby forging personnel identities and evading trajectory tracking. Traditional fixed threshold detection is prone to missed detections and false positives due to factors such as ambient lighting and equipment noise. This example supports flexible configuration of adaptive optimization cycles according to scenario requirements. For scenarios with high pedestrian traffic and high frequency of disturbance, such as urban roads and transportation hubs, a 7-day optimization cycle can be set, while for scenarios with slow pedestrian traffic, such as smart parks and office buildings, the optimization cycle can be adjusted to 15 days. It can also be adapted to extended application scenarios such as confidential factory areas, criminal investigation video gait evidence collection, outdoor inspection and monitoring in mining areas, and smart home gait recognition. All of these scenarios have inherent equipment noise and the risk of new hidden disturbance attacks from iterative updates, and this set of adaptive disturbance amplification detection logic can be reused in all of them.

[0075] Initial data settings:

[0076] Gait data dimensions: 3D gait feature data were selected from security scenarios ( The feature dimensions correspond to "stride length, stride frequency, and center of gravity height";

[0077] Preset optimization period: 7 days (to adapt to the dynamic adjustment needs of security scenarios).

[0078] The orthogonal matrix Q satisfies (3rd order identity matrix).

[0079] diagonal matrix , .

[0080] Magnification satisfies .

[0081] Step 201: Collect 1000 cumulative normal gait data and 200 cumulative abnormal disturbance samples in the security scenario according to the set period. After feature extraction, the maximum value of the inherent noise L2 norm of the cumulative normal gait data is 0.4, and the minimum disturbance L2 norm of the cumulative abnormal disturbance samples is 0.1.

[0082] Furthermore, based on the maximum L2 norm of 0.4 for normal noise, an increase of 20% is applied to calculate the optimal detection threshold. .

[0083] Based on the condition that the L2 norm after minimum perturbation amplification is greater than the optimal detection threshold, i.e. At the same time, according to the requirement that "the L2 norm of the amplified normal noise is less than the optimal detection threshold, i.e.: .

[0084] Select the optimal amplification parameters based on the overall adaptability to real-world scenarios. Its satisfaction And magnification times.

[0085] Furthermore, construct a diagonal matrix. , Its inverse matrix is, .

[0086] Generate orthogonal matrices :

[0087] Based on the feature matrix of normal gait data, a 3rd-order orthogonal matrix is ​​generated through QR decomposition. .

[0088] The optimal encoder matrix is ​​obtained by multiplying the orthogonal matrix, the diagonal matrix, and the transpose of the orthogonal matrix. The optimal decoder matrix is ​​then obtained by determining the inverse of the optimal encoder matrix, as shown below: ; .

[0089] Furthermore, the optimal detection threshold obtained above is... Optimal amplification parameters Optimal encoder matrix and the optimal decoder matrix This is stored as the initial condition for subsequent detection.

[0090] Step 202: Collect 3D initial gait feature data of a normal user, input the initial gait feature data into the optimal encoder matrix to obtain the initial ciphertext data. The specific formula includes: , .

[0091] Furthermore, the initial ciphertext data is transmitted to the optimal decoder matrix, and the initial ciphertext data received at the decoder is injected with the following hidden perturbation. (Injecting only 0.1 in the step size dimension), then obtaining the tampered ciphertext data, specifically: , .

[0092] Furthermore, the optimal decoder matrix is ​​invoked. The tampered ciphertext data is decoded to obtain the corrupted plaintext data, as shown below: .

[0093] In this example, the amplified perturbation is determined by the following formula: Furthermore, the amplification factor of the amplified perturbation is verified using the following formula: .

[0094] In this example, the optimal amplification parameter The theoretical upper limit of the global perturbation amplification factor is Since orthogonal matrices perform linear mapping transformations on multi-dimensional vectors, the perturbation energy is concentrated in the core feature dimension of step size, causing a significant increase in the amplitude of single-dimensional perturbation. Therefore, the calculation result of the perturbation amplitude amplification factor in single-dimensional perturbation is 25.

[0095] The formula for constraining the magnification of this scheme This example focuses on the overall L2 norm of the vector, rather than the magnitude of a single dimension. Only the step size dimension is perturbed; the original hidden perturbation exists only in one dimension. After matrix transformation, all perturbation energy converges to this dimension. Only the amplification factor of a single dimension exceeds the theoretical upper limit, while the overall L2 norm amplification factor of the vector still meets the constraint range. The core criterion for scheme detection and judgment is whether the overall L2 norm of the amplified perturbation exceeds the optimal detection threshold. It is only necessary to ensure that the vector norm meets the amplification factor constraint; there is no need to restrict the magnitude of a single dimension.

[0096] Step 203: Determine the amplified disturbance L2 norm, Amplify the disturbance Comparison of L2 norm and optimal detection threshold: because Therefore, we have: .

[0097] Based on this result, it can be determined that there is a hidden perturbation in the tampered encrypted data, triggering an alarm in the gait recognition system. Then, the obtained hidden perturbation is added to the accumulated abnormal perturbation sample to obtain an updated abnormal perturbation sample.

[0098] In this example, the following data was collected cumulatively during the 7-day testing period:

[0099] 200 new normal gait data points are added, and the maximum value of its inherent noise L2 norm becomes 0.45; 50 new cumulative abnormal perturbation samples are added, and the minimum perturbation L2 norm becomes 0.08; after reaching the preset period, a new round of step 201 is triggered. Based on the inherent noise L2 norm of this round and an increase of 20%, the updated optimal detection threshold is obtained, and the optimal amplification parameters, optimal encoder matrix and optimal decoder matrix are updated simultaneously.

[0100] In the new cycle, the updated optimal amplification parameter, optimal encoding matrix and optimal decoder matrix obtained from the above calculation are successively identified as the optimal detection threshold, optimal amplification parameter, optimal encoder matrix and optimal decoder matrix, and then the above steps 202-203 are repeated.

[0101] In summary, this invention provides a method and apparatus for detecting occult perturbations in gait identity recognition. The method performs full-dimensional optimization on accumulated data at a set period, and the output optimal parameters and matrices are adapted to the noise and perturbation features of the current scene, ensuring detection accuracy from the source and avoiding detection failures caused by changes in data distribution and iterative perturbation patterns in static frameworks. Furthermore, relying on encoder and decoder matrices constructed from orthogonal and diagonal matrices, it achieves directional perturbation amplification of core gait feature dimensions, transforming minute occult perturbations imperceptible to human vision and conventional detection into quantifiable amplified perturbations, while simultaneously ensuring normal... The gait data is losslessly encoded and decoded without affecting the system's original authentication function. Furthermore, the detection threshold and amplification parameters are dynamically set based on actual scene data, accurately distinguishing between normal noise and abnormal disturbances, effectively reducing the probability of false detection and missed detection. The parameters are also adapted to different application scenarios, improving the generalization of the solution. By classifying and updating the accumulated normal gait data and accumulated abnormal disturbance samples through detection results, the latest real data is provided for the next round of optimization, forming a positive closed loop of optimization, detection, data accumulation, and further optimization. This allows the detection capability to continuously iterate as the scene changes and the disturbance evolves, solving the problem of lagging feature library updates in traditional solutions.

[0102] Based on the same inventive concept, this invention provides a concealed disturbance detection device for gait identity recognition. Since the principle of this device in solving the technical problem is similar to that of the concealed disturbance detection method for gait identity recognition, the implementation of this device can refer to the implementation of the method, and the repeated parts will not be described again.

[0103] like Figure 2 As shown, the device includes a first obtaining unit 201, an amplification unit 202, and a judgment unit 203.

[0104] The first acquisition unit 201 is used to perform full-dimensional adaptive optimization on the accumulated normal gait data and accumulated abnormal disturbance samples in the acquired target scene according to a set period, so as to obtain the optimal detection threshold, optimal amplification parameters, optimal encoder matrix and optimal decoder matrix.

[0105] Amplification unit 202 is used to acquire initial gait feature data of a normal user in a target scene, input the initial gait feature data into the optimal encoder matrix to obtain initial ciphertext data; acquire tampered ciphertext data injected with hidden perturbations, input the tampered ciphertext data into the optimal decoder matrix to obtain amplified perturbations; wherein, the tampered ciphertext data includes the initial ciphertext data and the hidden perturbations; the hidden perturbations are amplified to obtain the amplified perturbations;

[0106] The judgment unit 203 is used to determine the L2 norm of the amplified perturbation. When the L2 norm of the amplified perturbation is greater than the optimal detection threshold, it is determined that there is a hidden perturbation in the tampered ciphertext data, and the hidden perturbation does not match the perturbation pattern included in the accumulated abnormal perturbation samples.

[0107] It should be understood that the units included in the above-described gait identity recognition concealment disturbance detection device are only a logical division based on the functions implemented by the device. In practical applications, the above units can be superimposed or separated. Furthermore, the functions implemented by the gait identity recognition concealment disturbance detection device provided in this embodiment correspond one-to-one with the gait identity recognition concealment disturbance detection method provided in the above embodiments. The more detailed processing flow implemented by this device has been described in detail in the above-described method embodiment one, and will not be described in detail here.

[0108] Another embodiment of the present invention provides a computer device, the computer device including: a processor and a scene database; the scene database is used to store computer program code, the computer program code including computer instructions; when the processor executes the computer instructions, the electronic device executes each step of the concealed perturbation detection method for gait identity recognition shown in the above method embodiment.

[0109] Another embodiment of the present invention provides a computer-readable storage medium storing computer instructions that, when executed on a computer device, cause the computer device to perform the various steps of the covert perturbation detection method for gait identity recognition shown in the above-described method embodiments.

[0110] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for detecting concealed perturbations in gait identity recognition, characterized in that, include: According to the set period, the accumulated normal gait data and accumulated abnormal disturbance samples in the target scene are subjected to full-dimensional adaptive optimization to obtain the optimal detection threshold, optimal amplification parameters, optimal encoder matrix and optimal decoder matrix. The initial gait feature data of a normal user in the target scenario is obtained, and the initial gait feature data is input into the optimal encoder matrix to obtain the initial ciphertext data; The tampered ciphertext data injected with a hidden perturbation is obtained, and the tampered ciphertext data is input into the optimal decoder matrix to obtain the amplified perturbation; wherein, the tampered ciphertext data includes the initial ciphertext data and the hidden perturbation; the hidden perturbation is amplified to obtain the amplified perturbation; Determine the L2 norm of the amplified perturbation. When the L2 norm of the amplified perturbation is greater than the optimal detection threshold, it is determined that there is a hidden perturbation in the tampered ciphertext data, and the hidden perturbation does not match the perturbation pattern included in the accumulated abnormal perturbation samples. The step involves performing full-dimensional adaptive optimization on the accumulated normal gait data and accumulated abnormal perturbation samples in the acquired target scene according to a set period, specifically including: The accumulated normal gait data comes from different user groups, collection environments, and device states; the accumulated abnormal perturbation samples include all discovered perturbation patterns; Based on the inherent noise level of the accumulated normal gait data, the maximum value of the noise L2 norm is determined, and the maximum value is increased by a preset ratio to obtain the optimal detection threshold. The optimal amplification parameter is determined based on all hidden perturbations included in the accumulated abnormal perturbation sample and the optimal detection threshold; wherein the optimal amplification parameter is greater than zero and less than 1, the L2 norm of the amplified perturbation is greater than the optimal detection threshold, and the L2 norm of the noise is less than the optimal detection threshold. The optimal encoder matrix is ​​determined by the product of an orthogonal matrix, a diagonal matrix, and the transpose of the orthogonal matrix; the first diagonal element of the diagonal matrix is ​​the optimal amplification parameter. The feature matrix of the accumulated normal gait data is decomposed using QR decomposition to obtain an orthogonal matrix that satisfies the set conditions. The optimal encoder matrix, the optimal decoder matrix, and the set conditions are as follows: in, Represents the optimal encoder matrix. Represents an orthogonal matrix. Represents a diagonal matrix. This represents the transpose of an orthogonal matrix. Represents the optimal decoder matrix. Represents a diagonal matrix The inverse matrix, Represents the identity matrix. This represents the optimal amplification parameter.

2. The method as described in claim 1, characterized in that, The dimension of the hidden perturbation is consistent with the dimension of the initial gait feature data; The amplification factor of the amplified disturbance satisfies: in, This indicates that the disturbance has been amplified. This indicates a concealed disturbance. This represents the optimal amplification parameter.

3. The method as described in claim 1, characterized in that, The step of inputting the tampered ciphertext data into the optimal decoder matrix to obtain the amplified perturbation specifically includes: The tampered ciphertext data is input into the optimal decoder matrix to obtain corrupted plaintext data represented by amplified perturbation and initial gait feature data. The corrupted plaintext data is shown below: in, This indicates that the plaintext data has been corrupted. This represents the initial gait feature data. This indicates that the disturbance has been amplified. Represents the optimal decoder matrix. This indicates a hidden disturbance.

4. The method as described in claim 1, characterized in that, After determining the L2 norm of the amplified perturbation, the method further includes: When the L2 norm of the amplified perturbation is less than or equal to the optimal detection threshold, it is determined that there is no hidden perturbation in the tampered ciphertext data. The initial gait feature data corresponding to the initial ciphertext data is added to the accumulated normal gait data to obtain updated normal gait data. When it is determined that there is a hidden perturbation in the tampered ciphertext data, the hidden perturbation is added to the accumulated abnormal perturbation sample to obtain an updated abnormal perturbation sample.

5. A concealed disturbance detection device for gait identity recognition, characterized in that, include: The first acquisition unit is used to perform full-dimensional adaptive optimization on the accumulated normal gait data and accumulated abnormal disturbance samples in the acquired target scene according to a set period, so as to obtain the optimal detection threshold, optimal amplification parameters, optimal encoder matrix and optimal decoder matrix. The amplification unit is used to acquire the initial gait feature data of a normal user in the target scene, and input the initial gait feature data into the optimal encoder matrix to obtain the initial ciphertext data; The tampered ciphertext data injected with a hidden perturbation is obtained, and the tampered ciphertext data is input into the optimal decoder matrix to obtain the amplified perturbation; wherein, the tampered ciphertext data includes the initial ciphertext data and the hidden perturbation; the hidden perturbation is amplified to obtain the amplified perturbation; The judgment unit is used to determine the L2 norm of the amplified perturbation. When the L2 norm of the amplified perturbation is greater than the optimal detection threshold, it is determined that there is a hidden perturbation in the tampered ciphertext data, and the hidden perturbation does not match the perturbation pattern included in the accumulated abnormal perturbation samples. The first obtaining unit is specifically used for: the accumulated normal gait data comes from different user groups, collection environments, and device states; the accumulated abnormal perturbation samples include all discovered perturbation patterns; Based on the inherent noise level of the accumulated normal gait data, the maximum value of the noise L2 norm is determined, and the maximum value is increased by a preset ratio to obtain the optimal detection threshold. The optimal amplification parameter is determined based on all hidden perturbations included in the accumulated abnormal perturbation sample and the optimal detection threshold; wherein the optimal amplification parameter is greater than zero and less than 1, the L2 norm of the amplified perturbation is greater than the optimal detection threshold, and the L2 norm of the noise is less than the optimal detection threshold. The optimal encoder matrix is ​​determined by the product of an orthogonal matrix, a diagonal matrix, and the transpose of the orthogonal matrix; the first diagonal element of the diagonal matrix is ​​the optimal amplification parameter. The feature matrix of the accumulated normal gait data is decomposed using QR decomposition to obtain an orthogonal matrix that satisfies the set conditions. The optimal encoder matrix, the optimal decoder matrix, and the set conditions are as follows: in, Represents the optimal encoder matrix. Represents an orthogonal matrix. Represents a diagonal matrix. This represents the transpose of an orthogonal matrix. Represents the optimal decoder matrix. Represents a diagonal matrix The inverse matrix, Represents the identity matrix. This represents the optimal amplification parameter.

6. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the covert perturbation detection method for gait identity recognition as described in any one of claims 1-4.

7. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the covert perturbation detection method for gait identity recognition as described in any one of claims 1-4.

Images