A multi-activity domain incremental learning-oriented continuous gait authentication method

Abstract

The application discloses a continuous gait identity authentication method for multi-activity domain incremental learning, and proposes a new identity authentication framework. First, the framework effectively decouples identity information and activity information by adopting a two-way encoder architecture, thereby enhancing the stability of identity representation under cross-activity domains. Second, an asymmetric memory management mechanism is introduced, and a differentiated feature representation storage and replay strategy is adopted to alleviate the catastrophic forgetting problem. Finally, an identity-aware contrast learning strategy is combined to guide the model to learn consistent cross-domain identity features, thereby improving authentication performance and anti-forgetting ability. A multi-objective optimization strategy is adopted to enable the model to maintain the authentication performance of the learned activity domain while having excellent adaptation ability to new activities. The method effectively alleviates the "catastrophic forgetting" problem, efficiently adapts to new activity scenarios while maintaining high authentication ability in the learned activity, and provides a practical solution for identity authentication in practical applications.

Description

Technical Field

[0001] This invention belongs to the field of identity authentication technology based on behavioral biometrics, specifically involving a continuous gait identity authentication method oriented towards incremental learning across multiple activity domains. The proposed method is implemented through a dual-encoder architecture, an asymmetric memory management mechanism, and an identity-aware contrastive learning strategy. Background Technology

[0002] Continuous gait authentication, as an important branch of behavioral biometric authentication technology, has received widespread attention from academia and industry in recent years. Unlike traditional one-time authentication methods, continuous gait authentication achieves implicit and imperceptible authentication throughout the entire session by continuously monitoring the user's walking patterns, effectively preventing unauthorized device access. With the widespread adoption of inertial sensor technology in mobile smart devices, gait data acquisition based on accelerometers and gyroscopes has become more efficient and convenient. In recent years, deep learning technology has driven the rapid development of the continuous gait authentication field. Deep learning models can effectively extract and learn gait features to achieve individual authentication. However, applying it to real-world scenarios still faces many challenges, especially the performance degradation problem in multi-activity scenarios.

[0003] In practical applications, users' daily activities include various movement patterns such as walking on flat ground, going up and down stairs. Gait characteristics differ significantly across different activities, making it difficult for models trained on a single activity domain to generalize to other activity types. More critically, as the types of user activities continue to expand, existing continuous gait authentication systems generally face the "catastrophic forgetting" problem. That is, when learning new activity patterns, the model significantly loses its ability to authenticate previously learned activities, leading to a decline in authentication performance and thus limiting the stability and reliability of the system in diverse practical applications.

[0004] While existing technologies offer various solutions to address the challenges of multiple activity domains, they all suffer from serious fundamental flaws. Traditional deep learning methods such as joint training, transfer learning, and multi-task learning require the pre-collection of data from all possible activity domains. This not only renders the system completely incapable of adapting to unknown activity types but also results in extremely high model building costs. When learning new gait activity patterns, simple periodic retraining strategies can partially alleviate the forgetting problem, but the resulting continuous computational burden and massive data storage requirements not only violate the privacy protection principles of mobile devices but also far exceed the actual carrying capacity of terminal devices, hindering the practical application of continuous gait authentication technology.

[0005] In continuous gait authentication scenarios, domain incremental learning offers a more practical solution. Unlike traditional incremental learning, which focuses on the addition of new user categories, domain incremental learning is specifically designed for scenarios where "identity categories are fixed while activity domains continuously increase," better matching the application characteristics of continuous gait authentication. As user activity types continuously expand, the model needs to adapt to new activity domains while maintaining the authentication performance of learned activities, posing a significant challenge to the model's continuous learning ability. Among numerous domain incremental strategies, compared with regularization-based or parameter isolation methods, replay mechanisms not only better handle significant distribution differences between different activity domains but also avoid model scaling with the number of activity domains, making them more suitable for resource-constrained mobile devices. By storing and replaying historical samples, replay mechanisms can effectively mitigate catastrophic forgetting problems and do not cause model scaling with the number of activity domains, making them particularly suitable for resource-constrained mobile devices. However, existing replay methods generally adopt a uniform memory management strategy, ignoring the asymmetry of the distribution of legitimate / illegal user features in continuous gait authentication tasks. This makes it difficult to optimize authentication performance with limited storage capacity and fails to meet the stringent requirements of high-reliability authentication. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, the continuous gait identity recognition method based on incremental learning across multiple activity domains provided by this invention solves the problem of "catastrophic forgetting" that occurs in existing gait authentication methods when user activity patterns are expanded, achieving efficient adaptation under new activities and high-performance maintenance under historical activities.

[0007] To achieve the aforementioned objectives, the present invention employs the following technical solution: a continuous gait authentication method oriented towards multi-activity domain incremental learning, comprising the following steps:

[0008] S1. Collect gait samples from multiple activity scenarios and construct a dataset;

[0009] The labels for the gait samples in the dataset include user labels and activity labels;

[0010] S2. Construct an asymmetric memory replay and identity-aware contrastive learning framework based on a dual-path feature extraction architecture, including a domain encoder, an identity encoder, and an asymmetric memory management module;

[0011] S3. Using the dataset, perform domain incremental learning on the asymmetric memory replay and identity perception contrast learning framework to obtain a gait identity authentication model.

[0012] During the domain incremental learning process, the activity features of the input gait data are extracted by the domain encoder, and the identity features of the input gait data are extracted by the identity encoder. Based on the extracted identity features, the asymmetric memory management module maintains independent memory pools for legitimate and illegitimate users respectively, and retains and replays the extracted historical identity features through differentiated storage and sampling strategies. Furthermore, the identity features extracted by the identity encoder and the asymmetric memory management module are used for identity-aware comparative learning to enhance the cross-domain activity consistency and user distinguishability of the features.

[0013] S4. Input the gait data to be authenticated into the gait identity authentication model and output the corresponding identity information.

[0014] Furthermore, the asymmetric memory management module includes a memory pool of finite size, a memory update unit, and a replay feature sampling unit;

[0015] The memory pool M is used to store identity features from previous activity domains, and when new activity domain data is input, it is used to jointly train the asymmetric memory management module by combining the current domain data and historical identity features in the memory pool; the memory pool M includes legitimate user memories M + With illegal user memory M - It stores the identity characteristics of legitimate users and illegitimate users respectively; among them, the legitimate user's memory M + A unified multi-domain management approach is adopted, while unauthorized user memories are managed independently in each domain.

[0016] The memory update unit is used to update the historical identity features stored in the memory pool using a quality assessment method;

[0017] The replay feature sampling unit is used to dynamically adjust the historical identity features sampled from each active domain according to the domain weight in each domain incremental session.

[0018] Furthermore, the method by which the memory update unit updates the historical identity features stored in the memory pool is specifically as follows:

[0019] S31. Construct a global prototype of a legitimate user. + and illegal user specific domain prototype In this system, legitimate users maintain a global prototype, while illegitimate users maintain a separate prototype for each activity domain.

[0020] S32. Calculate the quality score for each identity feature in the memory pool using a quality evaluation function; where, for the identity features of legitimate users, the global prototype p of legitimate users is used. + The assessment indicates that unauthorized users are using unauthorized user-specific domain prototypes. Evaluate;

[0021] S33. Based on the quality scores of each identity characteristic, select the highest-scoring s. k Each identity feature is retained in the memory of the activity domain k, thereby updating the historical identity features stored in the memory pool;

[0022] At the same time, the global prototype p of legitimate users + and illegal user specific domain prototype Update to the average value of the historical identity features retained in the corresponding memory buffer.

[0023] Further, in step S31, the legitimate user global prototype p + and illegal user specific domain prototype They are represented as follows:

[0024]

[0025] In the formula, h represents the identity feature, and M + Indicates legitimate user memory, This represents the invalid user memory corresponding to activity domain k;

[0026] In step S32, the quality score of the identity feature of the i-th sample is:

[0027]

[0028] In the formula, h i p represents the identity feature of the i-th sample. When sample i is a legitimate user sample, p represents the global prototype of the legitimate user. + When sample i is an illegal user sample, p represents the illegal user's specific domain prototype. The superscript T indicates the transpose operator.

[0029] Furthermore, the method for adjusting the number of identity features sampled by the replay feature sampling unit is as follows:

[0030] Calculate the domain weights of legitimate users in the memory pool respectively. Domain weight of unauthorized users according to and Dynamically allocate the number of feature representation samples of legitimate users in activity domain k The number of samples representing the characteristics of illegal users.

[0031] The number of samples based on the feature representations of legitimate and illegitimate users in each activity domain. and Based on the importance of the identity feature level, fine-grained identity feature sampling is performed in the activity domain k to obtain the most suitable combination of historical identity features in the activity domain k.

[0032] Furthermore, the domain weight of the legitimate user Domain weight of unauthorized users They are represented as follows:

[0033]

[0034]

[0035] In the formula, This represents the valid user memory in the activity domain k. p represents an illegal user memory in activity domain k. + Represents the global prototype of a legitimate user. This represents a specific domain prototype for an unauthorized user. This represents the specific domain prototype of the illegal user in the previous learning cycle, h represents the identity feature, and Div(·) represents the average Euclidean distance calculation function.

[0036] The number of feature representations corresponding to legitimate and illegitimate users in the activity domain k. and They are represented as follows:

[0037]

[0038] In the formula, N + and N - The feature representations of the asymmetric memory management modules allocated to legitimate and illegitimate users represent the upper limit of the number of samples. and These represent the domain weights of legitimate and illegitimate users in activity domain j, respectively. The subscript j represents the activity domain index, and K represents the total number of activity domains.

[0039] The process of fine-grained identity feature sampling in the activity domain k is represented as follows:

[0040]

[0041] In the formula, and These represent the legitimate and illegitimate user identity features sampled in the activity domain k, respectively. and Let represent the number of feature samples corresponding to legitimate and illegitimate users in the activity domain k, respectively. and h id p represents the identity features in the i-th sample. +This represents the global prototype of a legitimate user, Sampling(·) represents random sampling based on weights, WeightedProb(·) represents a function that converts identity features into sampling probabilities, and CosSim(·) represents the cosine similarity calculation function.

[0042] Furthermore, in step S3, during the incremental learning process, the training loss function L... total for:

[0043] L total =L cls +λ1L con +λ2L domain

[0044] In the formula, L cls L con and L domain Let λ1 and λ2 represent the identity authentication classification loss, identity perception contrastive learning loss, and domain adaptation loss, respectively, and let λ1 and λ2 represent the weights used to adjust the identity perception contrastive learning loss and domain adaptation loss in the total loss.

[0045] Furthermore, the identity-aware contrastive learning loss L con for:

[0046]

[0047] B = B current ∪S + ∪S -

[0048] P(i)={t∈B∣y t =y i and d t =d i ,t≠i}

[0049] In the formula, B represents the batch data constructed from the identity features in the current domain and the identity features of legitimate and illegitimate users sampled from the memory pool, and p(i) represents the set of identity features corresponding to the batch data B. This represents the identity features extracted from the i-th sample using the identity encoder. This indicates that all identity features in the identity feature set are being considered. Let τ represent the identity features extracted from the t-th sample using the ID card encoder, τ represent the temperature parameter, and Bi represent the features extracted from the batch excluding τi. All identity features outside of B are represented, sim(·,·) represents the cosine similarity function, and B current S represents the currently obtained identity feature representation. + and S -Let y represent the identity features of legitimate and illegitimate users sampled from the asymmetric memory module, respectively. t and y i The identity labels of the t-th and i-th samples respectively, d t and d i Let i and t represent the active domain labels of the t-th and i-th samples, respectively. The subscripts i and t represent different sample indices, and the subscript p represents the sample index that forms a positive sample pair with sample i.

[0050] Furthermore, the identity authentication classification loss L cls for:

[0051]

[0052] In the formula, N represents the size of the batch data, and σ(·) represents the sigmoid activation function. Let y represent the identity features extracted from the i-th sample using the identity encoder. i Let f(·) represent the identity label of the i-th sample, and f(·) represent the gait identity authentication model.

[0053] Furthermore, the domain adaptation loss L domain for:

[0054]

[0055] In the formula, C k This represents the activity feature center of the current activity domain k. This represents the prototype of the active domain j stored in the asymmetric memory management module, and n represents the number of available historical active domains.

[0056] The beneficial effects of this invention are as follows:

[0057] To address the problem of multi-activity domain incremental learning in continuous gait identity authentication, this invention proposes an asymmetric memory replay and identity-aware contrastive learning framework, which effectively alleviates the "catastrophic forgetting" problem. Specific advantages include:

[0058] (1) In this invention, the dual encoder architecture extracts user identity features through a three-layer convolutional network, while the domain encoder uses a two-layer convolutional network to capture activity-specific information. The former is used for identity discrimination and comparison loss, and the latter is used for domain center update and domain adaptation loss, thus achieving effective decoupling of identity features and activity domain features.

[0059] (2) The asymmetric memory management module in this invention uses a dual asymmetric mechanism to ensure that the most valuable identity feature representations are stored in memory through quality assessment, while the adaptive feature sampling strategy ensures that the most suitable combination of historical identity feature representations is used in each training. The two work together to implement differentiated storage and sampling strategies for legal and illegal user feature representations in the feature space, enabling AMRIC to efficiently cope with the challenges of multi-activity domain incremental learning with limited memory capacity.

[0060] (3) In this invention, an identity-aware contrastive learning strategy is used to map the same user features in different activity domains to similar feature spaces, enhancing cross-domain feature consistency while maintaining the consistency of identity feature representations for positive sample pairs. This contrastive learning mechanism not only enhances the feature differences between users but also significantly improves the feature stability of a single user when activities change, effectively mitigating the catastrophic forgetting problem in incremental activity domain scenarios.

[0061] (4) This invention employs a multi-objective joint optimization strategy to balance identity authentication, contrastive learning, and domain alignment objectives. This multi-objective optimization not only decouples identity and activity features but also effectively promotes synergistic enhancement among modules. The asymmetric memory management module prioritizes storing high-quality identity feature representations extracted by the identity encoder. These feature representations are used in subsequent identity-aware contrastive learning to construct cross-domain positive-negative pair feature representations, forming a positive feedback mechanism for continuous optimization. Identity-aware contrastive learning enhances the discriminativeness of identity feature representations and improves the representativeness of feature representations in the memory module. At the same time, the domain encoder learns activity features by minimizing domain adaptation loss, promoting cross-domain knowledge transfer and effectively mitigating catastrophic forgetting. Overall, the multi-objective optimization strategy enables AMRIC to maintain the authentication performance of learned activity domains while possessing excellent new activity adaptation capabilities, providing a robust and efficient solution for multi-activity domain incremental continuous gait authentication.

[0062] (5) The method of the present invention was fully experimentally verified on the public dataset USC-HAD and the self-built dataset CDUT-AG. The results show that the proposed method can maintain high authentication performance for learned activities without completely retraining the model when new activity types appear in sequence, effectively alleviating the "catastrophic forgetting" problem and verifying the effectiveness of the proposed method. Attached Figure Description

[0063] Figure 1 The flowchart of the continuous gait identity authentication method for incremental learning in multiple activity domains provided by this invention is shown.

[0064] Figure 2 The diagram shows the structural block of the Asymmetric Memory Replay and Identity Perception Comparative Learning Framework (AMRIC) provided by this invention.

[0065] Figure 3 This is a schematic diagram of the data increment implementation based on AMRIC provided by the present invention.

[0066] Figure 4 The comparison results provided by this invention are on the CDUT-AG dataset and the USC-HAD dataset.

[0067] Figure 5 The comparison results provided by this invention are for different numbers of active domains on the CDUT-AG dataset.

[0068] Figure 6 This invention provides comparative results on the USC-HAD dataset when the number of active domains increases.

[0069] Figure 7 The comparison results provided by this invention are for different activity domains on the CDUT-AG dataset and the USC-HAD dataset.

[0070] Figure 8 The results show the performance comparison of the ablation experiments of each core component of AMRIC on the CDUT-AG dataset and the USC-HAD dataset provided by this invention.

[0071] Figure 9 The present invention provides a comparison of different memory buffer sizes on the CDUT-AG and USC-HAD datasets.

[0072] Figure 10 The present invention provides a comparison of different memory allocation ratios between legitimate and illegitimate users on the CDUT-AG and USC-HAD datasets. Detailed Implementation

[0073] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0074] This invention provides a continuous gait authentication method based on incremental learning across multiple activity domains, such as... Figure 1 As shown, it includes the following steps:

[0075] S1. Collect gait samples from multiple activity scenarios and construct a dataset;

[0076] The labels for gait samples in the dataset include user labels and activity labels;

[0077] S2. Construct an asymmetric memory replay and identity-aware contrastive learning framework based on a dual-path feature extraction architecture, including a domain encoder, an identity encoder, and an asymmetric memory management module;

[0078] S3. Using the dataset, perform domain incremental learning on the asymmetric memory replay and identity perception contrastive learning framework to obtain the gait identity authentication model.

[0079] During the domain incremental learning process, the activity features of the input gait data are extracted by the domain encoder, and the identity features of the input gait data are extracted by the identity encoder. Based on the extracted identity features, the asymmetric memory management module maintains independent memory pools for legitimate and illegitimate users respectively, and retains and replays the extracted historical identity features through differentiated storage and sampling strategies. Furthermore, the identity features extracted by the identity encoder and the asymmetric memory management module are used for identity-aware comparative learning to enhance the cross-domain activity consistency and user distinguishability of the features.

[0080] S4. Input the gait data to be authenticated into the gait identity authentication model and output the corresponding identity information.

[0081] In this embodiment of the invention, in continuous gait authentication, the distribution of user gait features shifts due to changes in the activity scenario, making traditional methods difficult to adapt to long-term authentication tasks. To address this issue, in step S1 of this embodiment, the collected gait samples are formally defined as D = {D1, D2, ..., D...} K} represents the user's gait dataset under different activity types. Each activity domain contains high-dimensional time-series data acquired by inertial sensors. Where d is the number of sensor channels and T is the number of time steps for gait data. Each sample Associated identity tags in Representing the legitimate owner of the equipment, This represents an unauthorized, illegal user.

[0082] The continuous identity authentication method for incremental learning of multiple activity domains provided in this invention mainly solves the gait authentication problem under the continuous expansion of user activity scenarios. Its purpose is to achieve continuous and accurate authentication of legitimate users in scenarios where activity types are dynamically increasing, while avoiding catastrophic forgetting of the authentication performance of already learned activity domains.

[0083] Based on this, the Non-Repeating Memory Replay and Identity-Aware Contrast Learning Framework (AMRIC) constructed in step S2 of this embodiment adopts a domain incremental learning paradigm, organizing activity domains sequentially into incremental session sequences, with each session corresponding to a new activity type. During domain incremental learning, each new session can only access the training data of the current activity domain and a limited number of stored historical memory samples, and is not allowed to directly access the complete data of all historical activity domains. Simultaneously, in the evaluation phase after the training of each domain incremental session, the system needs to perform a comprehensive performance evaluation on the test samples of the current activity domain and all historical activity domains to verify the model's continuous authentication capability for various activity scenarios after learning in the new activity domain. Specifically, when a new activity domain k arrives, the model can only access the training data D of the current domain. k And a finite-size memory pool M, which contains data from the previous active domain {D1,D2,...,D...} k-1 The representative identity features selected in the} are represented by the memory pool M, which is subject to the constraint |M|≤C. M Limitation, C M This is a predefined upper limit for storage capacity.

[0084] In embodiments of the present invention, such as Figure 2 As shown, the asymmetric memory replay and identity-aware contrastive learning framework includes a domain encoder, an identity encoder, and an asymmetric memory management module. The framework adopts a dual-path feature extraction architecture to decouple identity and activity information. The identity encoder extracts user features through a three-layer convolutional network, while the domain encoder uses a two-layer convolutional network to capture specific activity information, providing clearer basic features for subsequent modules.

[0085] In embodiments of the present invention, such as Figure 2 The frame structure shown and Figure 3 The data increment process shown includes an asymmetric memory management module comprising a memory pool of finite size, a memory update unit, and a replay feature sampling unit.

[0086] Memory pool M stores identity features from previous activity domains and, when new activity domain data is input, combines the current domain data with historical identity features from the memory pool to jointly train the asymmetric memory management module; memory pool M includes legitimate user memories M + With illegal user memory M - It stores the identity characteristics of legitimate users and illegitimate users respectively; among them, the legitimate user's memory M + Adopt a unified multi-domain management approach Unauthorized user memories are managed independently by each domain. Where the size of memory pool M is |M|≤C M C M The upper limit of the predefined storage capacity;

[0087] The memory update unit is used to update the historical identity features stored in the memory pool using a quality assessment method;

[0088] The replay feature sampling unit is used to dynamically adjust the historical identity features sampled from each active domain according to the domain weight in each domain incremental session.

[0089] In this embodiment, the memory storage process of the memory pool adopts a differentiated strategy for legitimate and illegitimate users. During memory storage, the system first extracts the identity and activity features of gait samples through a dual-channel encoder. Furthermore, gait patterns are significantly affected by activity type; the same user exhibits large differences in different activity scenarios, while different users may have similar patterns in the same activity scenario. Therefore, the memory management of legitimate users needs to emphasize cross-domain consistency to maintain stable identity authentication capabilities, while the memory management of illegitimate users needs to maintain domain independence to accurately capture attack features in each scenario.

[0090] Specifically, for legitimate users, their identity characteristics are stored in a unified memory pool M. + In this context, the memory pool adopts a cross-activity domain organization approach, allowing legitimate user characteristics from different activity scenarios to jointly construct a global prototype p. + This ensures consistency in identity representation; however, for unauthorized users, their identity characteristics are stored separately in domain-independent memory pools M. - In this context, each active domain maintains an independent memory space and a corresponding domain-specific prototype. This allows the system to adapt to diverse gait patterns that may occur in different scenarios. Through this differentiated memory storage strategy, the system can achieve more effective knowledge retention and more accurate identity authentication with limited storage resources.

[0091] In this embodiment of the invention, in order to guide the quality management of the memory content in the memory pool, the method of updating the historical identity features stored in the memory pool through the memory update unit is specifically as follows:

[0092] S31. Construct a global prototype of a legitimate user. + and illegal user specific domain prototype In this system, legitimate users maintain a global prototype, while illegitimate users maintain a separate prototype for each activity domain.

[0093] Among them, the global prototype p of legitimate users + and illegal user specific domain prototype They are represented as follows:

[0094]

[0095] In the formula, h represents the identity feature, and M + Indicates legitimate user memory, This represents the invalid user memory corresponding to activity domain k;

[0096] S32. Calculate the quality score for each identity feature in the memory pool using a quality evaluation function; where, for the identity features of legitimate users, the global prototype p of legitimate users is used. + The assessment indicates that unauthorized users are using unauthorized user-specific domain prototypes. Evaluate;

[0097] The quality score of the identity feature of the i-th sample is:

[0098]

[0099] In the formula, h i p represents the identity feature of the i-th sample. When sample i is a legitimate user sample, p represents the global prototype of the legitimate user. + When sample i is an illegal user sample, p represents the illegal user's specific domain prototype. The superscript T indicates the transpose operator;

[0100] S33. Based on the quality scores of each identity characteristic, select the highest-scoring s. k Each identity feature is retained in the memory of the activity domain k, thereby updating the historical identity features stored in the memory pool;

[0101] At the same time, the global prototype p of legitimate users + and illegal user specific domain prototype Update to the average value of the historical identity features retained in the corresponding memory buffer;

[0102] The memory pool update process is represented as follows:

[0103] M k ←TopK(M k ∪H k ,q,s k )

[0104] In the formula, the TopK function determines whether a feature is retained in the memory pool based on the quality assessment q. The system first calculates the union M. k ∪H k Each identity feature h in i The quality rating is then selected, and the highest-rated s is chosen. k One remains in memory. H k s represents the identity characteristics of a legitimate or illegitimate user in activity domain k. k The memory capacity of field k

[0105] In this embodiment, the design of the aforementioned asymmetric storage mechanism is based on a deep understanding of gait authentication tasks. Gait patterns are significantly affected by activity types, and the same user may exhibit large differences in different activity scenarios. However, different users may have similar patterns in the same activity scenario. Therefore, the memory management of legitimate users needs to emphasize cross-domain consistency to maintain stable identity authentication capabilities, while the memory management of illegitimate users needs to maintain domain independence to accurately capture attack characteristics in each scenario. Through this differentiated memory storage strategy, the system can achieve more effective knowledge retention and more accurate identity authentication with limited storage resources.

[0106] In this embodiment of the invention, the aforementioned memory update method based on quality assessment ensures that the memory module stores representative identity feature representations. However, in continuous learning scenarios where the memory scale expands, directly using all stored identity feature representations leads to increased computational costs. Furthermore, when the number of identity feature representations in different activity domains is uneven, the model's authentication results tend to favor specific domains. Therefore, this invention designs an adaptive feature sampling unit that dynamically evaluates domain importance and the value of identity feature representations, extracting the most valuable subset from the memory to improve learning performance. For legitimate and illegitimate users, different domain importance calculation methods are used. The domain weight for legitimate users is based on the average similarity between all identity feature representations within that domain and the global prototype, encouraging the model to focus on the most representative identity features.

[0107] Specifically, in this embodiment, the method for adjusting the number of identity features sampled by the replay feature sampling unit is as follows:

[0108] Calculate the domain weights of legitimate users in the memory pool respectively. Domain weight of unauthorized users according to and Dynamically allocate the number of feature representation samples of legitimate users in activity domain k The number of samples representing the characteristics of illegal users.

[0109] The number of samples based on the feature representations of legitimate and illegitimate users in each activity domain. and Based on the importance of the identity feature level, fine-grained identity feature sampling is performed in the activity domain k to obtain the most suitable combination of historical identity features in the activity domain k.

[0110] Among them, the domain weight of legitimate users Domain weight of unauthorized users They are represented as follows:

[0111]

[0112]

[0113] In the formula, This represents the valid user memory in the activity domain k. p represents an illegal user memory in activity domain k. + Represents the global prototype of a legitimate user. This represents a specific domain prototype for an unauthorized user. Let represent the domain-specific prototype of the illegal user in the previous learning cycle, h represent the identity feature, and Div(·) represent the mean Euclidean distance calculation function, which measures the diversity of identity features within the domain. The degree of temporal change in the characteristics of the reaction domain;

[0114] The number of feature representations corresponding to legitimate and illegitimate users in the activity domain k. and They are represented as follows:

[0115]

[0116] In the formula, N + and N - The upper limit of the feature representation sampling number for the asymmetric memory management module allocated to legitimate users and illegitimate users, respectively (in actual training, typically 60% of the memory capacity is allocated to legitimate users and 40% to illegitimate users). and These represent the domain weights of legitimate and illegitimate users in activity domain j, respectively. The subscript j represents the activity domain index, and K represents the total number of activity domains.

[0117] The process of fine-grained identity feature sampling in the activity domain k is represented as follows:

[0118]

[0119] In the formula, and These represent the legitimate and illegitimate user identity features sampled in the activity domain k, respectively. and Let represent the number of feature samples corresponding to legitimate and illegitimate users in the activity domain k, respectively. and h id p represents the identity features in the i-th sample. + This represents the global prototype of a legitimate user, Sampling(·) represents random sampling based on weights, WeightedProb(·) represents a function that converts identity features into sampling probabilities, and CosSim(·) represents the cosine similarity calculation function.

[0120] In step S3 of this embodiment of the invention, during the incremental learning process, a multi-objective joint optimization strategy is adopted to integrate identity authentication loss, identity perception contrast loss and domain adaptation loss, guiding the identity encoder to learn feature representations that are related to user identity and consistent across activities, while the domain encoder learns activity-specific representations and achieves cross-domain alignment, thereby achieving a balance between learning new activities and maintaining historical knowledge.

[0121] Based on this, the training loss function L of the asymmetric memory replay and identity perception contrastive learning framework is... total for:

[0122] L total =L cls +λ1L con +λ2L domain

[0123] In the formula, L cls L con and L domain Let λ1 and λ2 represent the identity authentication loss, identity-aware contrastive learning loss, and domain adaptation loss, respectively, and let λ1 and λ2 represent the weights used to adjust the identity-aware contrastive learning loss and domain adaptation loss in the total loss.

[0124] In this embodiment, to address the challenge of identity feature consistency in domain incremental learning, an identity-aware contrastive learning strategy is designed. This strategy maximizes the feature similarity of the same user within the same activity domain by strictly defining positive and negative sample pairs, while minimizing the feature similarity between different users, thereby enhancing the cross-domain discriminative power of identity features. Specifically, batch data B is constructed by combining the feature representation of the current batch with the feature representations of legitimate and illegitimate users sampled from the memory pool. A set of positive feature representations P(i) is defined, and an identity authentication classification loss L is constructed. con for:

[0125]

[0126] B = B current ∪S + ∪S -

[0127] P(i)={t∈B∣y t =y i and d t =d i ,t≠i}

[0128] In the formula, B represents the batch data constructed from the identity features in the current domain and the identity features of legitimate and illegitimate users sampled from the memory pool, and p(i) represents the set of identity features corresponding to the batch data B. This represents the identity features extracted from the i-th sample using the identity encoder. This indicates that all identity features in the identity feature set are being considered. Let τ represent the identity features extracted from the t-th sample using the ID card encoder, τ represent the temperature parameter, and Bi represent the features extracted from the batch excluding τi. All identity features outside of B are represented, sim(·,·) represents the cosine similarity function, and B current S represents the currently obtained identity feature representation. + and S - Let y represent the identity features of legitimate and illegitimate users sampled from the asymmetric memory module, respectively. t and y i The identity labels of the t-th and i-th samples respectively, d t and d i Let i and t represent the active domain labels of the t-th and i-th samples, respectively. The subscripts i and t represent different sample indices, and the subscript p represents the sample index that forms a positive sample pair with sample i.

[0129] Furthermore, to facilitate the model's learning of domain-invariant identity features, the domain identifier for remembering identity features is uniformly set to the current activity domain. This approach allows the model to focus on distinguishing user identities rather than activity types, implicitly enhancing the feature consistency of the same user across different activity domains. In this way, the features of the same user in different activity domains are mapped to similar feature spaces while maintaining consistency in the definition of the identity feature representation. This identity-aware contrastive learning mechanism not only enhances the feature differences between users but also significantly improves the feature stability of a single user when activities change, effectively mitigating the catastrophic forgetting problem in scenarios with incremental activity domains.

[0130] In this embodiment, an identity authentication classification loss L is constructed. cls for:

[0131]

[0132] In the formula, N represents the size of the batch data, and σ(·) represents the sigmoid activation function. Let y represent the identity features extracted from the i-th sample using the identity encoder. i Let f(·) represent the identity label of the i-th sample, and f(·) represent the gait identity authentication model.

[0133] In this embodiment, the domain adaptation loss L is constructed. domain for:

[0134]

[0135] In the formula, C k This represents the activity feature center of the current activity domain k. This represents the prototype of the active domain j stored in the asymmetric memory management module, and n represents the number of available historical active domains.

[0136] In this embodiment, the multi-objective optimization not only decouples identity and activity features but also effectively promotes synergistic enhancement among modules. The asymmetric memory management module prioritizes storing high-quality identity feature representations extracted by the identity encoder. These feature representations are used in subsequent identity-aware contrastive learning to construct cross-domain positive-negative pair feature representations, forming a positive feedback mechanism for continuous optimization. Identity-aware contrastive learning enhances the discriminativeness of identity feature representations and improves the representativeness of feature representations in the memory module. Simultaneously, the domain encoder learns activity features by minimizing domain adaptation loss, promoting cross-domain knowledge transfer and effectively mitigating catastrophic forgetting. Overall, the multi-objective optimization strategy enables AMRIC to maintain the authentication performance of learned activity domains while possessing excellent new activity adaptation capabilities, providing a robust and efficient solution for incremental continuous gait authentication across multiple activity domains.

[0137] This invention provides an example of verifying the effectiveness of the above-described continuous gait authentication method.

[0138] In this embodiment, the proposed method was compared and analyzed with several other methods on the USC-HAD and CDUT-AG datasets, including Fine-tuning, Online Elastic Weight Consolidation (oEWC), Learning without Forgetting (LwF), Experience Replay (ER), Dark Experience Replay (DER++), Averaged Gradient Episodic Memory (A-GEM), and Error Sensitivity Modulation (ESM-ER). Three complementary evaluation metrics were used in the experiments: Equal Error Rate (EER) to measure single-domain authentication performance, Average Equal Error Rate (AE) to evaluate overall authentication performance across multiple activities, and Forgetting Metric (FM) to reflect knowledge retention capability.

[0139] The USC-HAD dataset contains inertial sensor data from 14 subjects (7 males and 7 females) during various activities. Data acquisition used a MotionNode inertial sensor (sampling frequency 100Hz) worn on the right anterior hip. Each subject performed five repetitions of each activity on different dates and in different environments. To focus on gait-related behaviors, we selected five activities with significant periodicity: Walking forward (WF), Walking left (WL), Walking right (WR), Walking upstairs (WU), and Walking downstairs (WD). These activities cover the most common activity patterns in the daily lives of mobile device users and are an ideal test set for evaluating authentication performance across activity domains.

[0140] The CDUT-AG dataset is a dataset collected for research on multi-activity identity authentication and gait recognition, containing data from 60 healthy participants aged 20-30 years. Data acquisition was conducted using sensor insoles, each insole integrating an inertial sensor (BMI160, including a three-axis accelerometer and a three-axis gyroscope) and 10 pressure sensors. Each participant performed three representative activities: level walking, walking upstairs, and walking downstairs. Gait data for level walking was collected in an 8-meter-long corridor, with each participant walking back and forth for approximately 2.5 minutes; walking up and downstairs was conducted in the stairwell, each session lasting approximately 15 seconds, with each activity repeated 10 times. The gait characteristics of different walking activities of the same individual showed significant differences, making it suitable as cross-activity domain test data.

[0141] In this embodiment, a differentiated preprocessing strategy was adopted to address the differences between the CDUT-AG and USC-HAD datasets. The CDUT-AG dataset incorporates time warp, amplitude perturbation, and random noise enhancement to simulate sensor errors and behavioral fluctuations, thereby improving model robustness. The USC-HAD dataset, however, already possesses natural diversity across multiple sessions and environments, so its original data was used. To realistically simulate the user behavior expansion process, the experiment employed a sequential domain incremental setup and open-set testing scheme: 10 configurations were constructed based on CDUT-AG (randomly selecting 1 legitimate user, 49 illegitimate users, and 10 attackers each time), and 3 configurations were constructed based on USC-HAD (randomly selecting 1 legitimate user, 10 illegitimate users, and 3 attackers each time). Legitimate user samples were divided into training and testing in a 7:3 ratio. The experiment used only inertial data, and the final performance was evaluated using multiple rounds of average assessment of the AE and FM metrics to ensure reliability. AE stands for Average Equal Error Rate, representing the average authentication capability across all learned activity domains after completing incremental learning in all domains; FM is the forgetting metric, evaluating the model's forgetting of previously learned tasks. Table 1 shows the typical experimental configuration and sample partitioning for the two datasets.

[0142] Table 1: Sample distribution of each activity domain in the USC-HAD and CDUT-AG datasets

[0143]

[0144] Detailed experimental results are shown in Tables 2 and 3.

[0145] Table 2: Comparison of results of different methods on the USC-HAD dataset

[0146]

[0147] Table 3: Comparison of results of different methods on the CDUT-AG dataset

[0148]

[0149]

[0150] From Table 2, Table 3 and Figure 4As can be seen, our method demonstrates significant advantages on both datasets. Specifically, our method significantly outperforms all comparable methods in both mean equal error rate (AE) and forgetting metric (FM), fully demonstrating its dual advantages in overall authentication accuracy and resilience to forgetting. On the CDUT-AG dataset, AMRIC achieves an AE of 1.17% and an FM of 1.24%, which are 0.43% and 0.32% lower than the suboptimal methods ER (AE 1.60%) and A-GEM (FM 1.56%), respectively. On the USC-HAD dataset, AMRIC achieves an AE of 1.52% and an FM of 1.16%, which are 1.52% and 1.00% lower than the suboptimal methods A-GEM (AE 3.04%) and DER++ (FM 2.16%), respectively. As the number of active domains increases from three in CDUT-AG to five in USC-HAD, the performance advantage of AMRIC expands significantly, indicating its good scalability when dealing with multiple active domains.

[0151] like Figure 5 and Figure 6 As shown, we demonstrate the performance of all methods as the number of active domains increases progressively, on the CDUT-AG dataset, through... Figure 5 (a) It can be observed that AMRIC maintains a stable performance advantage as the number of domains increases, achieving an average isoerror rate of 1.17% after completing incremental learning of all domains. Figure 5 As shown in the forgetting metric curve (b), AMRIC exhibits a continuous decreasing trend, eventually reaching 1.25%, significantly outperforming other methods. On the USC-HAD dataset, the experimental results show significant performance fluctuations. Figure 6 (a) It is evident that when the stair-climbing activity was introduced in Domain 4, the performance of all methods declined significantly. The mean equal error rate of AMRIC increased from 0.57% in Domain 3 to 4.73%; the performance of the ESM-ER method declined even more significantly, from 1.36% to 6.91%. However, in the Domain 5 (stair-climbing activity) stage, AMRIC showed a unique advantage, with the mean equal error rate decreasing to 1.52%, while other methods such as ESM-ER remained at a higher level. Figure 6 (b) shows that AMRIC's forgetting metric (1.16%) is also superior to other methods (such as DER++'s 2.16%). Experimental results demonstrate the consistent performance of our method across different datasets and scenarios of varying difficulty, validating its effectiveness and robustness as an incremental gait authentication method.

[0152] To analyze model performance from a more granular activity domain perspective, we present the equal error rate (EER) performance of each method on each individual activity domain after completing incremental learning across all activity domains, such as... Figure 7 As shown in the figure, on the CDUT-AG dataset, AMRIC performs excellently in all three activity domains, with EERs of 1.42% for "walking on flat ground," 1.90% for "going down stairs," and 0.19% for "going up stairs." In the up-and-down-stairs activities, AMRIC is 0.12 and 0.41 percentage points lower than the suboptimal method ESM-ER of 0.31% and 2.31%, respectively. Notably, AMRIC's EER variance across all activities is only 0.52, significantly lower than methods such as LwF (1.30) and A-GEM (0.77), demonstrating superior stability across activity domains. On the USC-HAD dataset, AMRIC also performs exceptionally well, especially in the "walking left" (WL) and "walking right" (WR) activities, with EERs of 0.62% and 1.86%, respectively, which are 0.11 and 1.63 percentage points lower than the suboptimal methods A-GEM and oEWC, respectively. In the "Wait-and-See" (WU) activity, AMRIC's EER (1.82%) was 0.49 percentage points lower than the second-best method, oEWC (2.31%). Overall, AMRIC's EER variance was only 0.24, significantly lower than comparable methods such as DER++ (2.76) and LwF (3.74). In particular, compared to oEWC (9.1490) and A-GEM (6.4304), AMRIC's variance was reduced by more than 96%, highlighting the cross-activity domain consistency of our method.

[0153] In this embodiment of the invention, experiments were also conducted on the proposed method under different activity sequences to verify the effectiveness and advancement of our method in domain incremental continuous identity authentication, providing reliable technical support for practical systems. The results are shown in Tables 4 and 5. On the CDUT-AG dataset, in six combinations of three-activity sequences, our method maintained a leading position in almost all activity sequences, with an average AE value of 0.73%, significantly outperforming existing methods such as A-GEM (1.02%) and ER (1.35%). On the more challenging USC-HAD dataset, in three representative five-activity sequences, our method achieved an average AE value of 1.87%, far lower than LwF (3.12%) and ESM-ER (3.18%). The results demonstrate that our method exhibits excellent generalization ability, maintaining a stable low error rate across different activity sequences. Furthermore, our method achieves an optimized balance between authentication accuracy and memory retention, enabling both key metrics to reach high levels simultaneously.

[0154] Table 4: Performance comparison of different activity orders on the CDUT-AG dataset (AE% / FM%)

[0155]

[0156] Table 5: Performance comparison of different activity sequences on the USC-HAD dataset (AE% / FM%)

[0157]

[0158] Furthermore, in this embodiment of the invention, ablation experiments were conducted on the proposed method to comprehensively evaluate the effectiveness of each functional component of the proposed framework and the impact of memory buffer parameter configuration on model performance. To comprehensively evaluate the contribution of each component to the proposed framework, we constructed five model variants for system comparison: (1) Removing the memory management module (w / o MM), training only with data from the current activity domain, verifying the role of the memory mechanism in preventing catastrophic forgetting; (2) Removing the dual encoder (w / o DE), using a single encoder to process identity and activity information simultaneously, failing to achieve feature decoupling and domain adaptation; (3) Removing the identity-aware contrastive learning strategy (w / o IACL), removing contrastive loss, retaining only classification and domain adaptation losses; (4) Randomized memory management (w / RM), using a random sample selection strategy instead of the asymmetric memory management strategy; and (5) The proposed full method was used as a benchmark model for comparison. The results are as follows: Figure 8 As shown.

[0159] The complete approach (AMRIC) achieves state-of-the-art performance on both datasets. Looking at the ablation effects of different components, removing the identity-aware contrastive learning strategy leads to a significant performance drop on the CDUT-AG dataset, indicating its significant role in enhancing feature consistency. On the USC-HAD dataset, which has more activity types, the absence of the memory management module causes an AE surge to 6.15%, demonstrating that an effective memory management module is crucial for knowledge retention. Furthermore, the dual-channel encoder shows consistent importance on both datasets; its removal increases the AE to 0.89% and 4.07%, respectively, indicating that the coupling of identity and activity features severely impacts the model's generalization ability. More notably, replacing the asymmetric memory management module with a random sample selection strategy results in a severe performance drop on the USC-HAD dataset, with AE and FM increasing to 10.71% and 13.01%, respectively. These results demonstrate that the synergistic effect among components is crucial for the stability of multi-activity domain incremental learning.

[0160] To further investigate the impact of memory buffer configuration on system performance, we systematically analyzed two key parameters: the size of the memory buffer and the memory allocation ratio between legitimate and illegitimate users. We conducted tests with various memory buffer size configurations on the CDUT-AG and USC-HAD datasets. Figure 9As shown, the CDUT-AG dataset achieves best performance with a memory buffer size of 50 (AE = 0.40%, FM = 0.55%), while the USC-HAD dataset achieves best performance with a buffer size of 100 (AE = 0.82%, FM = 1.01%). Figure 10 This study demonstrates the impact of the legitimate / illegal user memory allocation ratio on system performance. Despite the different activity domain sizes of the two datasets, optimal performance is achieved at a legitimate / illegal user memory allocation ratio of approximately 3:2 in both datasets. Specifically, the CDUT-AG dataset achieves optimal performance with a legitimate / illegal user memory allocation ratio of 30 / 20, while the USC-HAD dataset shows the best ratio at 60 / 40. The combined results indicate that maintaining a legitimate / illegal user memory allocation ratio of approximately 3:2 is crucial for achieving optimal performance, providing important guidance for memory resource allocation in practical deployments.

[0161] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0162] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A continuous gait authentication method oriented towards incremental learning across multiple activity domains, characterized in that, Includes the following steps: S1. Collect gait samples from multiple activity scenarios and construct a dataset; The labels for the gait samples in the dataset include user labels and activity labels; S2. Construct an asymmetric memory replay and identity-aware contrastive learning framework based on a dual-path feature extraction architecture, including a domain encoder, an identity encoder, and an asymmetric memory management module; S3. Using the dataset, perform domain incremental learning on the asymmetric memory replay and identity perception contrast learning framework to obtain a gait identity authentication model. During the domain incremental learning process, the activity features of the input gait data are extracted by the domain encoder, the identity features of the input gait data are extracted by the identity encoder, and the asymmetric memory management module maintains independent memory pools for legitimate users and illegitimate users based on the extracted identity features. The extracted historical identity features are retained and replayed through differentiated storage and sampling strategies. Furthermore, it performs identity perception comparison learning on the identity features extracted by the identity encoder and asymmetric memory management module to enhance the cross-domain activity consistency and user distinguishability of the features; S4. Input the gait data to be authenticated into the gait identity authentication model and output the corresponding identity information; The asymmetric memory management module includes a memory pool of finite size, a memory update unit, and a replay feature sampling unit; The memory pool M is used to store identity features from previous activity domains, and when new activity domain data is input, it combines the current domain data and historical identity features in the memory pool to jointly train the asymmetric memory management module; the memory pool M includes legitimate user memories. Memory of unauthorized users It stores the identity characteristics of legitimate users and illegitimate users respectively; among them, the legitimate user's memory... A unified multi-domain management approach is adopted, while unauthorized user memories are managed independently in each domain. The memory update unit is used to update the historical identity features stored in the memory pool using a quality assessment method; The replay feature sampling unit is used to dynamically adjust the historical identity features sampled from each active domain according to the domain weight in each domain incremental session.

2. The continuous gait authentication method based on incremental learning across multiple activity domains according to claim 1, characterized in that, The method by which the memory update unit updates the historical identity features stored in the memory pool is as follows: S31. Construct a global prototype for legitimate users. and illegal user specific domain prototype Among them, legitimate users maintain a global prototype, while illegitimate users maintain a separate prototype for each activity domain. S32. Calculate the quality score for each identity feature in the memory pool using a quality evaluation function; whereby the identity features of legitimate users are evaluated using the global prototype of legitimate users. The assessment indicates that unauthorized users are using unauthorized user-specific domain prototypes. Evaluate; S33. Based on the quality scores of each identity characteristic, select the highest score. Individual identity features are retained in the activity domain. In the memory, and then update the historical identity characteristics stored in the memory pool; At the same time, the global prototype of legitimate users will be... and illegal user specific domain prototype Update to the average value of the historical identity features retained in the corresponding memory buffer.

3. The continuous gait authentication method based on incremental learning across multiple activity domains according to claim 2, characterized in that, In step S31, the legitimate user global prototype and illegal user specific domain prototype They are represented as follows: In the formula, Indicates identity characteristics, Indicates legitimate user memory, This represents the invalid user memory corresponding to activity domain k; In step S32, the quality score of the identity feature of the i-th sample is: In the formula, p represents the identity feature of the i-th sample. When sample i is a legitimate user sample, p represents the global prototype of a legitimate user. When sample i is an illegal user sample, p represents the illegal user's specific domain prototype. The superscript T indicates the transpose operator.

4. The continuous gait authentication method based on incremental learning across multiple activity domains according to claim 1, characterized in that, The method for adjusting the number of sampled identity features in the replay feature sampling unit is as follows: Calculate the domain weights of legitimate users in the memory pool respectively. Domain weight of unauthorized users ,according to and Dynamically allocate the number of feature representation samples of legitimate users in activity domain k The number of samples representing the characteristics of illegal users. ; The number of samples based on the feature representations of legitimate and illegitimate users in each activity domain. and Based on the importance of the identity feature level, fine-grained identity feature sampling is performed in the activity domain k to obtain the most suitable combination of historical identity features in the activity domain k.

5. The continuous gait authentication method based on incremental learning across multiple activity domains according to claim 4, characterized in that, The domain weight of the legitimate user Domain weight of unauthorized users They are represented as follows: In the formula, This represents the valid user memory in the activity domain k. This represents an illegal user memory in activity domain k. Represents the global prototype of a legitimate user. This represents a specific domain prototype for an unauthorized user. This represents the domain-specific prototype of an illegal user from the previous learning cycle. Indicates identity characteristics, This represents the function for calculating the average Euclidean distance; The number of feature representations corresponding to legitimate and illegitimate users in the activity domain k. and They are represented as follows: In the formula, and The feature representations of the asymmetric memory management modules allocated to legitimate and illegitimate users represent the upper limit of the number of samples. and These represent the domain weights of legitimate and illegitimate users in activity domain j, respectively. The subscript j represents the activity domain index, and K represents the total number of activity domains. The process of fine-grained identity feature sampling in the activity domain k is represented as follows: In the formula, and These represent the legitimate and illegitimate user identity features sampled in the activity domain k, respectively. and Let represent the number of feature samples corresponding to legitimate and illegitimate users in the activity domain k, respectively. and , Let represent the identity features in the i-th sample. Represents the global prototype of a legitimate user. This indicates that random sampling is performed according to weights. This represents a function that converts identity features into sampling probabilities. This represents the function for calculating cosine similarity.

6. The continuous gait authentication method based on incremental learning across multiple activity domains according to claim 1, characterized in that, In step S3, during the incremental learning process, the loss function is trained. for: In the formula, , and These represent the loss for identity authentication classification, the loss for identity-aware contrastive learning, and the loss for domain adaptation, respectively. and These represent the weights used to adjust the identity perception contrastive learning loss and domain adaptation loss in the total loss, respectively.

7. The continuous gait authentication method based on incremental learning across multiple activity domains according to claim 6, characterized in that, The identity-perception contrastive learning loss for: In the formula, This represents batch data constructed from the identity features of users in the current domain and the identity features of legitimate and illegitimate users sampled from the memory pool. Represents batch data The corresponding set of positive identity features, This represents the identity features extracted from the i-th sample using the identity encoder. This indicates that all identity features in the identity feature set are being considered. This represents the identity features extracted from the t-th sample using the ID card encoder. Indicates temperature parameter, Indicates that, except for the batch All external identity characteristics represent, Represents the cosine similarity function. This represents the currently obtained identity feature representation. and These represent the identity features of legitimate and illegitimate users sampled from the asymmetric memory module, respectively. and The identity labels of the t-th and i-th samples respectively. and Let i and t represent the active domain labels of the t-th and i-th samples, respectively. The subscripts i and t represent different sample indices, and the subscript p represents the sample index that forms a positive sample pair with sample i.

8. The continuous gait authentication method based on incremental learning for multiple activity domains according to claim 6, characterized in that, The loss of identity authentication classification for: In the formula, N represents the size of the batch data. This represents the sigmoid activation function. This represents the identity features extracted from the i-th sample using the identity encoder. This represents the identity label of the i-th sample. This represents a gait-based identity authentication model.

9. The continuous gait authentication method based on incremental learning for multiple activity domains according to claim 6, characterized in that, The domain adaptation loss for: In the formula, This represents the activity feature center of the current activity domain k. This represents the prototype of the active domain j stored in the asymmetric memory management module, and n represents the number of available historical active domains.

Images