A method and system for protecting privacy by recognizing human activity in wearable devices

Abstract

This invention discloses a method and system for protecting privacy in wearable devices for human activity recognition, aiming to solve the problems of existing methods struggling to balance privacy and utility, and high resource consumption on the edge. The method first acquires multi-channel time-series signals from the wearable device and constructs fixed-length feature vectors. Based on a sensitive attribute inference model and the number of high-risk features, it divides the device into high-risk and low-risk feature sets. Only the high-risk feature sub-vectors are subjected to perturbation noise to generate perturbation latent representations, which are then uploaded to an edge server along with the low-risk feature set. The edge server generates the final feature vectors and assesses the risk of privacy leakage. An offline evaluation is used to construct a privacy-utility trade-off, dynamically adjusting the perturbation dimension and noise amplitude. This invention achieves targeted protection of high-risk features, maximizing the preservation of key recognition information, reducing edge computing power and communication overhead, adapting to the differentiated needs of different scenarios, and balancing privacy security and activity recognition performance.

Description

Technical Field

[0001] This invention relates to the field of privacy protection technology, and in particular to a method and system for protecting privacy by recognizing human activity in wearable devices. Background Technology

[0002] Wearable human activity recognition technology automatically identifies human behavior by collecting data from multi-channel sensors such as accelerometers, gyroscopes, and heart rates. It has been widely applied in scenarios such as motion monitoring, health management, and intelligent interaction. With the popularization of wearable devices, the high-dimensional time-series data they continuously collect is gradually becoming an important data foundation for supporting intelligent sensing and personalized services.

[0003] However, while such wearable timing signals are used for activity identification, they often implicitly contain sensitive information related to user identity, physiological characteristics, and lifestyle habits. Existing research indicates that even without explicit identification, attackers can still infer sensitive attributes such as a user's gender and height based on uploaded feature representations or model outputs, leading to privacy risks. This problem is particularly pronounced in cloud-edge-device collaborative architectures.

[0004] To address the privacy protection needs of wearable data, existing methods mainly fall into three categories:

[0005] First, there are static protection methods based on de-identification, encryption, or unified local differential privacy. These methods struggle to distinguish privacy risks across different feature dimensions and can significantly impact activity recognition performance. Second, there are edge training methods based on federated learning or adversarial representation learning. While these methods reduce centralized data uploads, they place high demands on terminal resources and system complexity. Third, there are privacy protection schemes that rely on fixed parameters or scenario configurations. These schemes lack a unified adjustment mechanism for the relationship between privacy strength and recognition performance, making them difficult to adapt to diverse application scenarios. Summary of the Invention

[0006] To address the problems of existing technologies, the present invention aims to provide a method and system for protecting privacy by recognizing human activity in wearable devices, which can meet the dual requirements of privacy security and system availability in practical applications.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for protecting privacy by recognizing human activity in a wearable device includes:

[0009] Acquire multi-channel time-series signals from wearable devices and construct fixed-length feature vectors for human activity recognition;

[0010] Based on the sensitive attribute inference model and the number of high-risk features, the fixed-length feature vector is divided to obtain a set of high-risk features and a set of low-risk features.

[0011] For high-risk feature subvectors in the high-risk feature set, perturbation noise is introduced to generate perturbation latent representations for human activity recognition;

[0012] Upload the potential representation of the perturbation and the set of low-risk features to the edge server;

[0013] On the edge server side, based on the perturbation latent representation and low-risk feature set, the final feature vector for human activity recognition is generated;

[0014] Privacy breach risk is assessed based on the final feature vector;

[0015] An offline assessment is conducted using factors such as privacy leakage risk, the number of high-risk features, the dimensions of the potential representation of perturbation, and the magnitude of perturbation noise to construct a privacy-utility trade-off.

[0016] Based on the privacy-utility trade-off, the dimensions of the perturbation potential representation and the magnitude of the perturbation noise are adjusted;

[0017] The steps for assessing privacy breach risk based on the final feature vector include:

[0018] The activity category is determined by using a human activity recognition model and the final feature vector.

[0019] The sensitive attribute prediction results are obtained by using a sensitive attribute inference model and the final feature vector.

[0020] The steps for constructing a privacy-utility tradeoff relationship through offline evaluation based on privacy leakage risk, the number of high-risk features, the dimension of the perturbation potential representation, and the magnitude of the perturbation noise include:

[0021] The results of activity category discrimination and sensitive attribute prediction are used as inputs for offline evaluation;

[0022] Under different combinations of the number of high-risk features, the dimension of the perturbation potential representation, and the amplitude of the perturbation noise, the accuracy of human activity recognition and the accuracy of sensitive attribute inference were repeatedly calculated.

[0023] Under the condition of meeting the utility threshold of human activity recognition, a privacy-utility trade-off relationship is constructed based on the accuracy of human activity recognition and the accuracy of sensitive attribute inference.

[0024] As a further improvement of the present invention, the step of acquiring multi-channel time-series signals collected by a wearable device and constructing a fixed-length feature vector for human activity recognition includes:

[0025] Acquire multi-channel timing signals from wearable devices;

[0026] Multi-channel time-series signals are segmented by sliding according to fixed window length and step size, and statistical features are extracted in each window to construct a fixed-length feature vector.

[0027] As a further improvement of the present invention, the step of dividing a fixed-length feature vector based on a sensitive attribute inference model and the number of high-risk features to obtain a high-risk feature set and a low-risk feature set includes:

[0028] A sensitive attribute inference model is used to evaluate the contribution of each feature dimension in the fixed-length feature vector to the prediction of sensitive attributes, and the data is aggregated on the sample dimension to obtain the privacy sensitivity score corresponding to each feature dimension.

[0029] The feature dimensions are sorted according to the size of the privacy sensitivity score;

[0030] The highest-scoring high-risk features are selected from a number of feature dimensions to form a high-risk feature set.

[0031] The remaining feature dimensions are used to construct a low-risk feature set.

[0032] As a further improvement of the present invention, the step of introducing perturbation noise into the high-risk feature sub-vectors in the high-risk feature set to generate a perturbation latent representation for human activity recognition includes:

[0033] For high-risk feature subvectors in the high-risk feature set, a compression coding module is used for encoding mapping, compressing them into a low-dimensional latent space, and perturbation noise is introduced into the low-dimensional latent space to generate perturbation latent representations for human activity recognition.

[0034] As a further improvement of the present invention, the step of uploading the perturbation latent representation and the low-risk feature set to the edge server includes:

[0035] The perturbation latent representation is concatenated with low-risk feature sub-vectors from the low-risk feature set to form the upload payload;

[0036] The payload is uploaded to the edge server via a communication link.

[0037] As a further improvement of the present invention, the step of generating the final feature vector for human activity recognition on the edge server side based on the perturbation latent representation and the low-risk feature set includes:

[0038] The potential representation of the disturbance and the low-risk feature vector are obtained from the uploaded payload.

[0039] High-risk features are reconstructed from the latent representation of perturbation using a generative reconstruction model;

[0040] The reconstructed high-risk feature vectors are concatenated with the low-risk feature vectors to generate the final feature vector for human activity recognition.

[0041] As a further improvement of the present invention, the present invention also includes the steps of: constructing a joint optimization objective that includes high-risk feature reconstruction constraints, human activity recognition and discrimination constraints and privacy suppression constraints, and jointly training the compression coding module and the generative reconstruction model.

[0042] The present invention also provides a human activity recognition privacy protection system for implementing the above-mentioned human activity recognition privacy protection method, comprising:

[0043] The sensing module is used to acquire multi-channel time-series signals collected by wearable devices and construct a fixed-length feature vector for human activity recognition.

[0044] The feature segmentation module is used to segment a fixed-length feature vector based on the sensitive attribute inference model and the number of high-risk features, to obtain a set of high-risk features and a set of low-risk features.

[0045] The perturbation latent representation generation module is used to introduce perturbation noise into the high-risk feature sub-vectors in the high-risk feature set to generate perturbation latent representations for human activity recognition.

[0046] The transmission module is used to upload the potential representation of the perturbation and the set of low-risk features to the edge server;

[0047] The final feature vector generation module is used on the edge server side to generate the final feature vector for human activity recognition based on the perturbation latent representation and low-risk feature set.

[0048] The evaluation module is used to assess the risk of privacy breaches based on the final feature vector.

[0049] The trade-off module is used to construct a privacy-utility trade-off relationship through offline evaluation based on privacy leakage risk, the number of high-risk features, the dimension of the perturbation potential representation, and the magnitude of the perturbation noise.

[0050] The optimization module is used to adjust the dimension of the perturbation potential representation and the magnitude of the perturbation noise based on the privacy-utility trade-off.

[0051] The beneficial effects of this invention are as follows: Based on a sensitive attribute inference model and a high-risk feature quantity threshold, this invention selectively divides fixed-length feature vectors, classifying features containing sensitive information such as user identity and physiological characteristics into a high-risk set, while retaining the original information of other low-risk features. Perturbation is applied only to high-risk features, avoiding the destruction of useful information from low-risk features. This effectively prevents privacy leaks while maximizing the preservation of key features required for human activity recognition, significantly reducing the negative impact of privacy protection on recognition performance.

[0052] The edge device only needs to complete feature segmentation and perturbation generation of high-risk feature sub-vectors, without performing complex full-data encryption operations. This significantly reduces the requirements for computing power, storage, and energy consumption of wearable devices, making it perfectly suited for application scenarios with limited terminal resources. Furthermore, by uploading only the perturbation latent representation and low-risk feature set, compared to the original high-dimensional time-series signal or full features, the amount of data transmitted in the end-to-edge communication link is effectively reduced, lowering communication overhead and the risk of data leakage.

[0053] By constructing a privacy-utility trade-off through offline evaluation, this approach overcomes the shortcomings of traditional solutions that rely on fixed parameters and cannot flexibly adapt to different scenarios. In actual deployment, the dimensions of the perturbation potential representation and the amplitude of the perturbation noise can be dynamically adjusted according to the differentiated requirements of privacy protection strength and recognition performance in different application scenarios (such as high privacy protection in health management scenarios and high recognition accuracy in motion monitoring scenarios), thereby achieving a personalized balance between privacy protection and task utility. Attached Figure Description

[0054] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0055] Figure 1 This is a flowchart of the human activity recognition privacy protection method described in this invention;

[0056] Figure 2 This is another flowchart of the human activity recognition privacy protection method described in this invention. Detailed Implementation

[0057] To make the technical problems solved by the present invention, the technical solutions adopted, and the technical effects achieved clearer, the technical solutions of the embodiments of the present invention will be further described in detail below. 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.

[0058] This invention provides a method for protecting privacy by recognizing human activity in wearable devices, such as... Figure 1 and Figure 2 As shown, it includes:

[0059] S1. Acquire multi-channel time-series signals from the wearable device at the edge, and construct a fixed-length feature vector for human activity recognition. The multi-channel time-series signals include sensor data such as acceleration and gyroscope readings, represented as... , , indicating the first Multi-channel timing signals collected by users. For the length of time, 6 represents the number of channels, corresponding to the three-axis accelerometer and three-axis gyroscope signals.

[0060] For multi-channel timing signals, use a fixed window length and step length Perform sliding segmentation, for the first segment... Time segments extracted from each window The feature processing is performed, and its expression is:

[0061]

[0062] Wherein, the window length is taken. Step length Adjacent windows overlap by 50%. Feature processing. The process includes: within each sliding window, extracting statistical features for each of the six sensor channels. These features include the mean, standard deviation, maximum value, and mean absolute deviation, resulting in four statistical features for each channel and a total of 24 statistical features across the six channels. Simultaneously, calculating inter-axis correlation coefficients for the three-axis signals from both the accelerometer and gyroscope sensors to characterize the correlation between signals along different axes within the same sensor. Each set of three-axis signals corresponds to three inter-axis correlation coefficients, resulting in a total of six correlation features across the two sets of sensors. Finally, concatenating the 24 statistical features with the six correlation features constructs a 30-dimensional fixed-length feature vector.

[0063]

[0064] S2. Based on the sensitive attribute inference model and the number of high-risk features, the fixed-length feature vector is divided to obtain a set of high-risk features and a set of low-risk features.

[0065] For example, a sensitive attribute inference model can be pre-trained using gender as a sensitive attribute. Sensitive attribute inference model The SHAP feature attribution method is used to evaluate the contribution of each feature dimension in a fixed-length feature vector to the prediction of sensitive attributes. 3D features in the 1st dimension The contribution on each sample is expressed as:

[0066] ;

[0067] in This indicates that, considering all possible permutations of each feature dimension of a fixed-length eigenvector, the th... The average marginal contribution of dimensional features to the prediction results of the sensitive attribute inference model. This indicates that under a certain characteristic arrangement, the first element is not included. The input representation of the features of dimension, This indicates adding the first [number] [item] to this permutation. The input representation after 3D features, To obtain the expectation of all permutations of features.

[0068] Then, aggregation is performed along the sample dimension, that is, the absolute values ​​of the contributions are averaged along the sample dimension to obtain the privacy sensitivity score corresponding to each feature dimension:

[0069]

[0070] in, The number of samples used for offline risk assessment. The larger the value, the more likely it is to be the first The stronger the overall contribution of a feature to the prediction of sensitive attributes, the higher the risk of privacy leakage. (Based on privacy sensitivity scores) Sort by highest to lowest, and select the highest number of features. The high-risk feature set consists of several feature dimensions. The remaining feature dimensions constitute a low-risk feature set. , .

[0071] S3. For the high-risk feature sub-vectors in the high-risk feature set, introduce perturbation noise to generate perturbation potential representations for human activity recognition.

[0072] In this process, only the high-risk feature set is considered. The high-risk feature vectors in the data are processed, and the fixed-length feature vector obtained in step S1 is denoted as . After step S2, the top 10 feature dimensions with the highest privacy sensitivity scores are selected to form a high-risk feature set, corresponding to high-risk feature sub-vectors. A compression encoding module is used to process high-risk feature subvectors. Encoding mapping is performed, compressing the data into a low-dimensional latent space. Perturbation noise is then introduced into this low-dimensional latent space to generate a perturbation latent representation for human activity recognition. The expression for this perturbation latent representation is as follows:

[0073]

[0074] in, This represents the latent mean function after mean shrinkage. A fixed variance parameter is used to control the amplitude of disturbance noise in the latent space. For the identity matrix, in this embodiment, we take... The generated perturbation latent representation has a dimension of .

[0075] Unlike high-risk feature vectors, low-risk feature vectors Without applying compression or perturbation noise, it maintains low-distortion passthrough for subsequent feature fusion and human activity recognition processing.

[0076] S4. Upload the potential representation of the perturbation and the set of low-risk features to the edge server.

[0077] First, perturb the latent representation Low-risk feature sub-vectors in the low-risk feature set The data is then concatenated to form the upload payload; subsequently, the upload payload is uploaded to the edge server via a communication link. The upload payload is represented as:

[0078] ;

[0079] Among them, symbols This represents a vector concatenation operation. Therefore, the uploaded payload maintains the same dimensions as the original fixed-length feature vector while only concatenating some of them. The high-risk feature dimensions were compressed and perturbed.

[0080] Each sample on the edge only needs to upload a fixed-length feature vector of length 30, without uploading the original multi-channel time-series signal. This effectively reduces the computational and communication overhead on the edge while meeting privacy requirements. The uploaded payload is sent to the edge server via a wireless communication link for subsequent generative reconstruction and human activity recognition processing.

[0081] S5. On the edge server side, based on the perturbation latent representation and low-risk feature set, the final feature vector for human activity recognition is generated.

[0082] Step S5 is completed on the edge server side, where the edge server receives the uploaded payload. Among them, the potential representation of perturbation The low-dimensional representations of the corresponding high- and low-risk feature vectors are used at the edge, employing a generative reconstruction model. Decoding and reconstructing the perturbation latent representation yields the reconstructed high-risk feature vector, expressed as follows:

[0083]

[0084] Generative Reconstruction Model A decoding network structure based on variational autoencoder (VAE) is adopted, which maps the latent space representation back to the high-risk feature space through fully connected layers. Subsequently, the reconstructed high-risk feature sub-vectors are... With low-risk feature vectors The features are then stitched and fused to form the final feature representation used for human activity recognition:

[0085]

[0086] The final feature representation, while maintaining the same dimension as the original fixed-length feature vector, weakens the information in high-risk features that can be used for sensitive attribute inference through generative reconstruction, providing a unified input for subsequent human activity recognition and privacy risk assessment.

[0087] S6. Assess privacy leakage risk based on the final feature vector, and then... The data is input into a human activity recognition model and a sensitive attribute inference model for inference, resulting in activity category discrimination results and sensitive attribute prediction results. The human activity recognition model... Sensitive attribute inference model Both employ lightweight one-dimensional convolutional neural network structures, and the human activity recognition model... Pre-training is performed based on labeled historical data before deployment.

[0088] The activity category determination result is expressed as follows:

[0089]

[0090] Indicates the first The activity category classification results corresponding to each sample are used by the human activity recognition model to distinguish between them. Similar to daily activities.

[0091] The results of the sensitive attribute prediction are expressed as follows:

[0092]

[0093] This represents the prediction results of sensitive attributes obtained based on the final feature representation. The sensitive attribute inference model is used to evaluate the inferability of gender attributes, and its prediction accuracy is used as a privacy leakage risk assessment indicator.

[0094] The training process, which involves collaboration between the end-user and edge servers, also includes the following steps: constructing a joint optimization objective that includes constraints for high-risk feature reconstruction, human activity recognition and discrimination, and privacy suppression, and jointly training the compression coding module and the generative reconstruction model.

[0095] The joint optimization objective is expressed as:

[0096]

[0097] During training, a joint optimization objective is used to optimize the compression coding module and the generative reconstruction model. Guided by task discrimination constraints and privacy suppression constraints, the generated final feature vector balances the effectiveness of human activity recognition with the protection of sensitive attributes.

[0098] In the formula, This is the reconstruction loss term, used to constrain the consistency between the reconstructed high-risk features and the original high-risk features; The regularization loss term is used to impose distribution constraints on the latent representation of high-risk features to stabilize controlled perturbations and match differential privacy analysis. This is used as a loss term for task discrimination to maintain the performance of human activity recognition; This is a privacy-suppressing loss term, used to weaken the inferability of sensitive attributes. and These are adjustable weighting parameters used to balance task utility with the strength of privacy protection. This is the regularization weight parameter, used to adjust the degree of influence of distribution constraints on the overall optimization objective.

[0099] S7. Conduct offline assessments based on privacy leakage risks, the number of high-risk features, the dimensions of perturbation potential representations, and the magnitude of perturbation noise to construct a privacy-utility trade-off relationship.

[0100] Using the activity category discrimination results and sensitive attribute prediction results from step S6 as inputs for offline evaluation, different combinations of privacy protection parameters are selected offline for comparative analysis; the privacy protection parameter combinations include the number of high-risk features. Perturbation latent representation dimension and disturbance noise amplitude For each set of parameters, the accuracy of the activity recognition model output is repeatedly calculated and used as a task utility metric. The accuracy of repeatedly calculating the output of the sensitive attribute inference model is used as a privacy risk indicator. Under the condition of meeting the threshold for human activity recognition utility, the task utility index corresponding to different parameter combinations is mapped to the privacy risk index as a two-dimensional evaluation point, and the lower limit constraint of the task utility index is met. Filter among the candidate points, where, The minimum acceptable accuracy threshold for human activity recognition is pre-set based on actual application requirements. Subsequently, following the principle of lower privacy risk and higher task utility, a candidate optimal boundary, i.e., the Pareto front, is determined, thereby establishing a privacy-utility trade-off relationship for deployment decisions.

[0101] S8. Based on the privacy-utility trade-off, adjust the dimension of the perturbation potential representation and the magnitude of the perturbation noise.

[0102] During the system deployment phase, based on the specific application scenario's requirements for recognition performance and privacy protection strength, corresponding parameter configurations are selected from the privacy-utility trade-off relationship to achieve adjustable deployment of privacy protection strength. For example, in medical monitoring scenarios, a "high privacy" configuration (a larger number of high-risk features and a higher perturbation noise amplitude) can be prioritized to reduce the inferability of sensitive attributes; in sports guidance scenarios, a "high utility" configuration (a moderate number of high-risk features and a lower perturbation noise amplitude) can be prioritized to maintain the accuracy of human activity recognition.

[0103] The privacy protection method for human activity recognition in this invention is implemented based on the PyTorch deep learning framework (version 1.12.2). The core network modules include a variational autoencoder (VAE) for modeling high-risk feature perturbations, a convolutional neural network (CNN) classifier for human activity recognition, and an identity inference CNN model for sensitive attribute inference. All of these models are implemented in the PyTorch environment. The system runs on Ubuntu 20.04. During training and inference, an NVIDIA RTX 3090 GPU (24GB VRAM) is used, along with an Intel Xeon-class CPU and 128GB of RAM to meet the computational requirements for joint training of multiple models and privacy assessment. The edge-side modules can be deployed on wearable or mobile terminal devices with limited computing power, while the edge-side modules are deployed on local servers or edge computing nodes.

[0104] In summary, the human activity recognition privacy protection method of the present invention has the following beneficial effects:

[0105] To address the challenges of easily inferring privacy information in wearable data within a cloud-edge-device collaborative environment, the limited computing and communication resources on the device side, and the difficulty in balancing privacy protection and recognition performance, this paper proposes a privacy processing mechanism based on feature-level risk perception and selective protection. By assessing the privacy risks of feature dimensions, privacy constraints are imposed and compressed only on high-risk feature sets, while low-risk feature sets are kept unbiased and pass-through. This effectively reduces unnecessary perturbations. Compared to traditional methods that uniformly add noise to all features or compress them overall, this approach significantly improves the usability of human activity recognition tasks while maintaining privacy protection capabilities.

[0106] This invention further introduces a latent compressed representation mechanism that satisfies differential privacy constraints at the edge side. This mechanism maps high-risk feature sub-vectors from a high-risk feature set to a low-dimensional latent space and applies formal privacy protection to the uploaded representation through perturbation noise injection. This effectively limits the inferability of sensitive attributes even under semi-honest threat models where communication link eavesdropping or edge node leakage is possible. Compared to privacy protection schemes that rely on fixed perturbation strength or lack theoretical constraints, this invention enables representation uploads with clear privacy boundaries under lightweight computational conditions at the edge side, enhancing the security and controllability of the system in practical deployment scenarios.

[0107] Furthermore, this invention introduces a generative reconstruction model and joint training mechanism on the edge server side to reconstruct high-risk features from the perturbation latent representation. These features are then fused with low-risk feature sub-vectors from the low-risk feature set to form the final feature vector. This allows the reconstructed features to effectively weaken sensitive attribute patterns while preserving human activity discrimination information. By offline evaluation of the number of high-risk features and privacy-related parameters (dimensionality of the perturbation latent representation, amplitude of perturbation noise), this invention establishes a trade-off between privacy protection strength and task utility, enabling the system to flexibly select appropriate privacy protection configurations according to different application scenarios. Experimental results show that on typical wearable datasets, this invention significantly reduces the re-identification ability of sensitive attributes such as gender and height while maintaining relatively stable activity recognition performance. It also exhibits good robustness and generalization ability when the training sample size changes, demonstrating high practical application and promotion value.

[0108] Based on the same inventive concept, this invention also provides a human activity recognition privacy protection system for wearable devices, used in the aforementioned human activity recognition privacy protection method, comprising: a perception module, a feature segmentation module, a perturbation latent representation generation module, a transmission module, a final feature vector generation module, an evaluation module, a trade-off module, and an optimization module. The perception module acquires multi-channel time-series signals collected by the wearable device to construct a fixed-length feature vector for human activity recognition; the feature segmentation module divides the fixed-length feature vector based on a sensitive attribute inference model and the number of high-risk features to obtain a high-risk feature set and a low-risk feature set; the perturbation latent representation generation module is used to analyze the high-risk features in the high-risk feature set... The system comprises several modules: a risk feature sub-vector, which introduces perturbation noise to generate a perturbation latent representation for human activity recognition; a transmission module for uploading the perturbation latent representation and a low-risk feature set to an edge server; a final feature vector generation module for generating a final feature vector for human activity recognition on the edge server based on the perturbation latent representation and the low-risk feature set; an evaluation module for assessing the privacy leakage risk based on the final feature vector; a trade-off module for constructing a privacy-utility trade-off relationship through offline evaluation based on privacy leakage risk, the number of high-risk features, the dimension of the perturbation latent representation, and the magnitude of the perturbation noise; and an optimization module for adjusting the dimension of the perturbation latent representation and the magnitude of the perturbation noise based on the privacy-utility trade-off relationship.

[0109] Please refer to the previous text for the specific implementation process of the human activity recognition privacy protection system, which will not be repeated here.

[0110] In this description, references to terms such as "an embodiment," "example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example.

[0111] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style of the specification is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0112] The technical principles of the present invention have been described above with reference to specific embodiments. These descriptions are merely for explaining the principles of the invention and should not be construed as limiting the scope of protection of the invention in any way. Based on this explanation, those skilled in the art can readily conceive of other specific embodiments of the invention without inventive effort, and these embodiments will all fall within the scope of protection of the present invention.

Claims

1. A method for protecting privacy by recognizing human activity in a wearable device, characterized in that, include: Acquire multi-channel time-series signals from wearable devices and construct fixed-length feature vectors for human activity recognition; Based on the sensitive attribute inference model and the number of high-risk features, the fixed-length feature vector is divided to obtain a set of high-risk features and a set of low-risk features. For high-risk feature subvectors in the high-risk feature set, perturbation noise is introduced to generate perturbation latent representations for human activity recognition; Upload the potential representation of the perturbation and the set of low-risk features to the edge server; On the edge server side, based on the perturbation latent representation and low-risk feature set, the final feature vector for human activity recognition is generated; Privacy breach risk is assessed based on the final feature vector; An offline assessment is conducted using factors such as privacy leakage risk, the number of high-risk features, the dimensions of the potential representation of perturbation, and the magnitude of perturbation noise to construct a privacy-utility trade-off. Based on the privacy-utility trade-off, the dimensions of the perturbation potential representation and the magnitude of the perturbation noise are adjusted; The steps for assessing privacy breach risk based on the final feature vector include: The activity category is determined by using a human activity recognition model and the final feature vector. The sensitive attribute prediction results are obtained by using a sensitive attribute inference model and the final feature vector. The steps for constructing a privacy-utility tradeoff relationship through offline evaluation based on privacy leakage risk, the number of high-risk features, the dimension of the perturbation potential representation, and the magnitude of the perturbation noise include: The results of activity category discrimination and sensitive attribute prediction are used as inputs for offline evaluation; Under different combinations of the number of high-risk features, the dimension of the perturbation potential representation, and the amplitude of the perturbation noise, the accuracy of human activity recognition and the accuracy of sensitive attribute inference were repeatedly calculated. Under the condition of meeting the utility threshold of human activity recognition, a privacy-utility trade-off relationship is constructed based on the accuracy of human activity recognition and the accuracy of sensitive attribute inference.

2. The method for protecting privacy by recognizing human activity according to claim 1, characterized in that, The step of acquiring multi-channel time-series signals collected by the wearable device and constructing a fixed-length feature vector for human activity recognition includes: Acquire multi-channel timing signals from wearable devices; Multi-channel time-series signals are segmented by sliding according to fixed window length and step size, and statistical features are extracted in each window to construct a fixed-length feature vector.

3. The method for protecting privacy by recognizing human activity according to claim 1, characterized in that, The step of dividing a fixed-length feature vector into a high-risk feature set and a low-risk feature set based on a sensitive attribute inference model and the number of high-risk features includes: A sensitive attribute inference model is used to evaluate the contribution of each feature dimension in the fixed-length feature vector to the prediction of sensitive attributes, and the data is aggregated on the sample dimension to obtain the privacy sensitivity score corresponding to each feature dimension. The feature dimensions are sorted according to the size of the privacy sensitivity score; The highest-scoring high-risk features are selected from a number of feature dimensions to form a high-risk feature set. The remaining feature dimensions are used to construct a low-risk feature set.

4. The method for protecting privacy by recognizing human activity according to claim 1, characterized in that, The step of introducing perturbation noise into the high-risk feature sub-vectors in the high-risk feature set to generate perturbation latent representations for human activity recognition includes: For high-risk feature subvectors in the high-risk feature set, a compression coding module is used for encoding mapping, compressing them into a low-dimensional latent space, and perturbation noise is introduced into the low-dimensional latent space to generate perturbation latent representations for human activity recognition.

5. The method for protecting privacy by recognizing human activity according to claim 4, characterized in that, The step of uploading the perturbation latent representation and the low-risk feature set to the edge server includes: The perturbation latent representation is concatenated with low-risk feature sub-vectors from the low-risk feature set to form the upload payload; The payload is uploaded to the edge server via a communication link.

6. The method for protecting privacy by recognizing human activity according to claim 5, characterized in that, The step of generating the final feature vector for human activity recognition on the edge server side, based on the perturbation latent representation and the low-risk feature set, includes: The potential representation of the disturbance and the low-risk feature vector are obtained from the uploaded payload. High-risk features are reconstructed from the latent representation of perturbation using a generative reconstruction model; The reconstructed high-risk feature vectors are concatenated with the low-risk feature vectors to generate the final feature vector for human activity recognition.

7. The method for protecting privacy by recognizing human activity according to claim 6, characterized in that, It also includes the following steps: constructing a joint optimization objective that includes constraints for high-risk feature reconstruction, human activity recognition and discrimination, and privacy suppression, and jointly training the compression coding module and the generative reconstruction model.

8. A human activity recognition privacy protection system for wearable devices, characterized in that, A method for implementing human activity recognition privacy protection as described in any one of claims 1 to 7, comprising: The sensing module is used to acquire multi-channel time-series signals collected by wearable devices and construct a fixed-length feature vector for human activity recognition. The feature segmentation module is used to divide a fixed-length feature vector based on the sensitive attribute inference model and the number of high-risk features, to obtain a set of high-risk features and a set of low-risk features. The perturbation latent representation generation module is used to introduce perturbation noise into the high-risk feature sub-vectors in the high-risk feature set to generate perturbation latent representations for human activity recognition. The transmission module is used to upload the potential representation of the perturbation and the set of low-risk features to the edge server; The final feature vector generation module is used on the edge server side to generate the final feature vector for human activity recognition based on the perturbation latent representation and low-risk feature set. The evaluation module is used to assess the risk of privacy breaches based on the final feature vector. The trade-off module is used to construct a privacy-utility trade-off relationship through offline evaluation based on privacy leakage risk, the number of high-risk features, the dimension of the perturbation potential representation, and the magnitude of the perturbation noise. The optimization module is used to adjust the dimension of the perturbation potential representation and the magnitude of the perturbation noise based on the privacy-utility trade-off.

Images