A privacy protection federated learning updating system and method for an intelligent terminal ecology

Abstract

This invention belongs to the field of distributed privacy computing and cross-domain authentication technology, specifically involving a privacy-preserving federated learning update system and method for the smart terminal ecosystem. The system includes the following steps: acquiring multimodal feature sequences and topological state parameters; decoupling heterogeneous protocols and aligning cross-domain feature spaces; generating adaptive privacy perturbation masks and gradient masking; routing to federated coordination nodes and dynamic aggregation; distributing the updated global authentication model to candidate takeover terminals; performing cross-domain feature inference to generate a confidence distribution; and triggering a session state snapshot transfer when the confidence distribution satisfies the dynamic takeover condition and the spatiotemporal continuity parameters between the source terminal and the candidate takeover terminal cross a preset migration boundary. This invention can solve the problems of rigid authentication across branded devices and privacy leakage.

Description

Technical Field

[0001] This invention belongs to the field of distributed privacy computing and cross-domain authentication technology, specifically relating to a privacy-preserving federated learning and update system and method for the smart terminal ecosystem. Background Technology

[0002] With the rapid development of the smart terminal ecosystem, multi-device collaboration has become a mainstream application scenario. Currently, it often relies on the same account system or brand-specific protocol to achieve mutual trust between devices. In terms of model updates and privacy protection, a basic federated learning framework is usually adopted. This involves training a global identity verification model by collecting local multimodal feature sequences and performing gradient routing based on cross-domain device topology state parameters, in order to achieve smooth handover of permissions and transfer of session state snapshots between terminals.

[0003] Problems with existing technology: However, in actual working conditions, when heterogeneous protocol terminals are connected in the scenario, traditional trust transfer mechanisms are prone to blind diffusion. Specifically, the differences in sensor sampling of different brands of devices lead to significant representational shifts in multimodal feature sequences. If the gradient is updated locally directly, it will cause the risk of leakage of identity-sensitive data. In addition, existing solutions lack joint measurement of device heterogeneity and local context risk entropy, resulting in rigid dynamic takeover condition judgment, significant verification delay when switching high-privilege operations, and difficulty in generating protocol-independent local identity representation vectors and ensuring cross-domain privacy and security. Summary of the Invention

[0004] The purpose of this invention is to provide a privacy-preserving federated learning update system and method for the smart terminal ecosystem, which can solve the problems of rigid cross-brand device verification and privacy leakage.

[0005] The specific technical solution adopted by this invention is as follows: A privacy-preserving federated learning update method for the smart terminal ecosystem includes the following steps: S1. Obtain the multimodal feature sequence and cross-domain device topology state parameters collected by the source terminal node, perform heterogeneous protocol decoupling and cross-domain feature space alignment on the multimodal feature sequence, and generate protocol-independent local identity representation vectors. S2. Calculate the feature residual between the local identity representation vector and the historical global identity benchmark model, construct a local gradient update vector based on the feature residual, and generate an adaptive privacy perturbation mask according to the device heterogeneity and the security level of the interaction event, and perform targeted masking on the local gradient update vector to obtain a hierarchical desensitized gradient vector. S3. The hierarchical desensitization gradient vector is routed to the federated coordination node. The federated coordination node calculates the dynamic aggregation weight based on the cross-domain trust decay factor, local data distribution offset and device brand heterogeneity penalty coefficient, performs gradient fusion operation, and outputs the global identity verification model for the next iteration cycle. S4. The updated global authentication model is distributed to the candidate takeover terminal, and cross-domain feature inference is performed to generate a confidence distribution. When the confidence distribution meets the dynamic takeover conditions and the spatiotemporal continuity parameter between the source terminal and the candidate takeover terminal crosses the preset migration boundary, the session state snapshot transfer is triggered.

[0006] According to another aspect of the present invention, the steps of heterogeneous protocol decoupling and cross-domain feature space alignment include: Extract the device protocol bias component and the general biometric component from the multimodal feature sequence to construct a heterogeneous feature decoupling matrix; The heterogeneous feature decoupling matrix is ​​mapped to a cross-domain shared latent space by regularized orthogonal projection. The local identity representation vector is then iteratively updated using the feature cluster centers within the cross-domain shared latent space to remove the representation bias caused by the sampling differences of sensors from different brands of terminals.

[0007] According to another aspect of the present invention, the logic for generating the adaptive privacy perturbation mask includes: Based on the security level of the interactive event, a basic permission gradient and a high-order permission gradient are divided. Low-variance directional noise injection is configured for the high-order permission gradient, and a structured perturbation mask is configured for the basic permission gradient. The noise injection intensity is nonlinearly positively correlated with the device heterogeneity, so that the hierarchical desensitization gradient vector retains the high-frequency behavioral evolution information required for cross-domain verification while satisfying the hierarchical differential privacy budget constraint.

[0008] According to another aspect of the present invention, the calculation of the dynamic aggregation weight includes constructing a joint cost function of a topology trust decay function, a data distribution offset metric, and a device brand heterogeneity penalty coefficient, and optimizing the joint cost function to determine the optimal contribution weight of each terminal node. When the local data distribution offset of any terminal node exceeds a preset deviation threshold, the optimal contribution weight is subjected to gradient pruning and channel isolation to suppress the parameter pollution of the global authentication model by heterogeneous distribution.

[0009] According to another aspect of the present invention, the cross-domain feature deduction and dynamic takeover condition determination include: Perform local forward propagation computation on the candidate takeover terminal side to generate a multidimensional confidence distribution vector; The risk entropy value of the current environment and the time-series inertial window of user behavior are analyzed in real time, and the takeover confidence threshold range is dynamically adjusted based on the risk entropy value. When the peak percentage of the multidimensional confidence distribution vector crosses the dynamically adjusted takeover confidence threshold and the entropy fluctuation range is within the preset stable range, the cross-domain feature deduction is determined to be successful.

[0010] According to another aspect of the present invention, the determination logic of the spatiotemporal continuity parameter includes: Construct a spatial proximity decay trajectory between the source terminal node and the candidate takeover terminal, and align the decay trajectory with the user behavior temporal inertial window through convolution; Calculate the rate of change of the derivative of the spatial proximity decay trajectory and the temporal overlap of the behavior characterized by convolution alignment; When the rate of change of the derivative is less than a preset smoothing limit and the temporal overlap of the behavior crosses a preset continuous limit, a seamless takeover instruction is generated.

[0011] According to another aspect of the present invention, a privacy-preserving federated learning update method for the smart terminal ecosystem further includes a non-registered subject intervention processing step: When a secondary biometric subject not associated with the primary account is identified in the multimodal feature sequence, the static physiological components and short-term interaction patterns of the secondary biometric subject are extracted to generate a temporary shadow identity vector. The temporary shadow identity vector is projected onto a preset low-dimensional permission subspace and bound to an independent session lifecycle and instruction whitelist. Within the lifecycle, it only responds to local gradient update requests corresponding to basic control instructions.

[0012] According to another aspect of the present invention, a privacy-preserving federated learning update system for the smart terminal ecosystem is also provided, comprising: The heterogeneous protocol bridging unit is configured to capture multimodal data streams from cross-brand terminal nodes, perform device protocol bias stripping and cross-domain feature space alignment, and output protocol-independent local identity representation vectors. The hierarchical gradient decoupling engine establishes a hardware-level isolated communication link with the heterogeneous protocol bridging unit. It has a built-in adaptive perturbation mask generator and a directional noise injector, which are used to perform component separation and privacy masking on the local update gradient according to the security level of the interaction event and the heterogeneity of the device, and output a hierarchical desensitized gradient summary. The federated dynamic aggregation gateway integrates a cross-domain trust routing table and a joint cost optimizer. The cross-domain trust routing table maintains the trust decay trajectory and brand heterogeneity mapping relationship of each node. The joint cost optimizer reconstructs the aggregation channel topology based on the local data distribution offset and heterogeneity penalty coefficient, performs multi-source gradient fusion, and broadcasts the global authentication kernel. The scenario-adaptive takeover controller, connected to the federated dynamic aggregation gateway, integrates a risk entropy assessment module and a spatiotemporal continuity matcher. When the risk entropy assessment module outputs a low-disturbance state and the spatiotemporal continuity matcher determines that the behavior sequence is coherent, it generates cross-node state transition control signaling and synchronizes the encrypted session context to the candidate takeover terminal.

[0013] According to another aspect of the present invention, the hierarchical gradient decoupling engine is deployed in the secure execution environment of the terminal device. The secure execution environment is configured with a gradient decoupling bus and a raw feature circuit breaker to block the direct pass-through path of the unde-sensitized raw multimodal sequence to the external network, and only allows the gradient summary after component separation and directional noise injection to be routed to the federated dynamic aggregation gateway through the gradient decoupling bus.

[0014] According to another aspect of the present invention, the scenario-adaptive takeover controller further includes a lightweight state packer and a near-field synchronization protocol stack, configured to compress the session context of the source terminal into an incremental state snapshot after the cross-node state flow control signaling is triggered. The incremental state snapshot includes desensitized gradient routing mask parameters and a dynamic takeover threshold identifier, and is asynchronously broadcast to candidate takeover terminals through a near-field direct connection protocol or a local area network multicast channel.

[0015] According to another aspect of the present invention, an electronic device is also provided, the electronic device including a memory and a processor; the memory is used to store a program; the processor executes the program to implement the method described in any one of the foregoing.

[0016] According to another aspect of the present invention, a computer-readable storage medium is also provided, the storage medium storing a computer program that, when executed by a processor, implements the method described in any one of the preceding embodiments.

[0017] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the method described in any one of the preceding embodiments.

[0018] The technical effects achieved by this invention are as follows: This invention separates the device protocol bias component and the general biometric component in the multimodal feature sequence by employing heterogeneous protocol decoupling and cross-domain feature space alignment. It also utilizes regularized orthogonal projection to construct a protocol-independent local identity representation vector, enabling the identity features of different brand terminals to still be mapped to a unified mathematical representation space even when their underlying sampling architecture and communication protocols are inconsistent. This eliminates the model convergence deviation caused by hardware heterogeneity in cross-brand mutual trust scenarios and realizes brand-barrier-free migration of identity verification in a multi-terminal ecosystem.

[0019] This invention introduces an adaptive privacy perturbation mask generation logic, splits the local update gradient into high and low permission dimensions based on the security level of the interaction event, and implements targeted noise injection by combining device heterogeneity and hierarchical differential privacy budget. This allows high-order identity-sensitive gradients to retain effective signal features under strict privacy constraints, while low-sensitivity interaction gradients are structurally masked. This avoids the global identity verification model from losing its discriminative ability due to excessive perturbation, while ensuring the on-demand allocation and fine-grained control of the privacy budget during cross-domain updates, effectively blocking gradient inversion attack paths.

[0020] This invention constructs a joint cost function that includes a topological trust decay function, a data distribution offset metric, and a device brand heterogeneity penalty coefficient. This function dynamically optimizes and isolates the aggregated channel weights within the federated coordination nodes, automatically reducing the weights of terminal nodes whose local data distribution deviates significantly from the baseline or whose trust trajectories decay. This suppresses the parameter pollution of the global model caused by heterogeneous distribution anomalies, thereby substantially improving the model's anti-interference capability during multi-source gradient fusion and ensuring the convergence stability of federated learning updates in complex network topology environments.

[0021] This invention deploys a context-adaptive takeover controller and combines risk entropy assessment with spatiotemporal continuity matching logic. Based on real-time analysis of environmental threat level and user behavior inertia, it dynamically adjusts the verification confidence threshold. When the spatial proximity decay trajectory and the overlap of behavioral time sequence meet the continuity boundary, it triggers incremental synchronization of state snapshots. This makes cross-device permission handover no longer rely on static strong verification links, but achieves a smooth transition through multi-dimensional context awareness, thereby significantly reducing the authentication latency and user interaction interruption frequency in multi-terminal switching scenarios. Attached Figure Description

[0022] Figure 1 This invention relates to a privacy-preserving federated learning update method process. Figure 1 ; Figure 2 This invention relates to a privacy-preserving federated learning update method process. Figure 2 ; Figure 3 This is a block diagram of the privacy-preserving federated learning update system architecture of the present invention. Detailed Implementation

[0023] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0024] It should be noted that the terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] According to an embodiment of the present invention, a method embodiment of a privacy-preserving federated learning update method for the smart terminal ecosystem is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0026] like Figure 1 As shown, a privacy-preserving federated learning update method for the smart terminal ecosystem includes the following steps: S1. Obtain the multimodal feature sequence and cross-domain device topology state parameters collected by the source terminal node, perform heterogeneous protocol decoupling and cross-domain feature space alignment on the multimodal feature sequence, and generate protocol-independent local identity representation vectors. S2. Calculate the feature residual between the local identity representation vector and the historical global identity benchmark model, construct the local gradient update vector based on the feature residual, and generate an adaptive privacy perturbation mask according to the device heterogeneity and the security level of the interaction event. Perform targeted masking on the local gradient update vector to obtain the hierarchical desensitized gradient vector. S3. The hierarchical desensitization gradient vector is routed to the federated coordination node. The federated coordination node calculates the dynamic aggregation weight based on the cross-domain trust decay factor, local data distribution offset and device brand heterogeneity penalty coefficient, performs gradient fusion operation, and outputs the global identity verification model for the next iteration cycle. S4. The updated global authentication model is distributed to the candidate takeover terminal. Cross-domain feature inference is performed to generate a confidence distribution. When the confidence distribution meets the dynamic takeover conditions and the spatiotemporal continuity parameter between the source terminal and the candidate takeover terminal crosses the preset migration boundary, the session state snapshot transfer is triggered.

[0027] The above steps are executed through a collaborative computing architecture deployed at the edge of the terminal device and the cloud federated server.

[0028] Further, in step S1, the source terminal node first collects the user's biometric signals and environmental interaction data in real time through its built-in heterogeneous sensor array, forming a multimodal feature sequence. Simultaneously, it scans the communication handshake messages of surrounding devices through the underlying network interface to parse the cross-domain device topology state parameters. During this process, a privacy-preserving federated learning update system (hereinafter referred to as the system) for the smart terminal ecosystem performs heterogeneous protocol decoupling and cross-domain feature space alignment on the multimodal feature sequence. By constructing a protocol-independent feature extraction network, it removes the sampling bias introduced by different hardware underlying drivers, thereby generating a protocol-independent local identity representation vector. Thus, the construction logic of the local identity representation vector is independent of specific device brands, laying the foundation for subsequent cross-brand terminal mutual trust.

[0029] Building upon this, step S2 further calculates the feature residuals between the local identity representation vector and the historical global identity benchmark model, and constructs a local gradient update vector based on these residuals. However, considering the differentiated privacy requirements in cross-device interactions, the system comprehensively evaluates the current device heterogeneity and the security level of interaction events to generate an adaptive privacy perturbation mask. The adaptive privacy perturbation mask performs a targeted masking operation on the local gradient update vector, applying structured noise to gradients in high-sensitivity dimensions while retaining a high signal-to-noise ratio in low-sensitivity dimensions, resulting in a hierarchical desensitized gradient vector. This on-demand noise injection mechanism ensures the security of high-order identity features while preventing the global model from losing its generalization ability due to excessive perturbation.

[0030] Subsequently, in step S3, the hierarchical desensitization gradient vector is routed to the federated coordination node via an encrypted communication channel. The node is equipped with a dynamic aggregation optimization engine. After receiving gradient vectors from multiple terminals, it comprehensively evaluates the cross-domain trust decay factor, local data distribution offset, and device brand heterogeneity penalty coefficient of each node, constructs a multi-dimensional joint cost function for iterative optimization, calculates the dynamic aggregation weights, performs weighted fusion operations, and outputs the global authentication model for the next iteration. By introducing a brand heterogeneity penalty coefficient, model drift caused by differences in sensor characteristics from different manufacturers is effectively suppressed.

[0031] Finally, in step S4, the updated global authentication model is distributed to the candidate takeover terminals, and cross-domain feature inference is performed to generate a confidence distribution. When the system determines that the confidence distribution meets the dynamic takeover conditions and the spatiotemporal continuity parameter monitored in real time crosses the preset migration boundary, the session state snapshot transfer is automatically triggered. This process relies on the joint matching of device spatial proximity trajectory and user behavior inertia, enabling the handover of permissions to be completed without the user's awareness, significantly reducing the verification latency during cross-terminal switching.

[0032] As an optional embodiment, the steps of heterogeneous protocol decoupling and cross-domain feature space alignment include: Extract the device protocol bias component and the general biometric component from the multimodal feature sequence, and construct a heterogeneous feature decoupling matrix; By using regularized orthogonal projection, the heterogeneous feature decoupling matrix is ​​mapped to the cross-domain shared latent space. The local identity representation vector is then iteratively updated using the feature clustering centers within the cross-domain shared latent space to remove the representation shift caused by the sampling differences between sensors from different brands of terminals.

[0033] Based on the above, when performing heterogeneous protocol decoupling and cross-domain feature space alignment, the system first uses a pre-trained modality separation network to separate the device protocol bias component and the general biometric component from the multimodal feature sequence, and then constructs a heterogeneous feature decoupling matrix accordingly.

[0034] Furthermore, to eliminate differences in sensor resolution, sampling frequency, and signal encoding format among different brand terminals, a regularized orthogonal projection algorithm is preferably used to map the matrix to a cross-domain shared latent space. Within this latent space, the system iteratively optimizes the local identity representation vector by maintaining dynamically updated feature cluster centers.

[0035] Furthermore, the regularization term introduces manifold alignment constraints during the projection process, which forces the features collected by different devices to be aligned to a unified representation manifold, fundamentally eliminating the feature space offset caused by hardware heterogeneity. As a result, even if the terminal device adopts a completely independent private communication protocol, its generated local identity representation vector can still maintain mathematical consistency on a global scale, ensuring feature comparability during cross-brand verification.

[0036] In one alternative embodiment, the cross-domain shared latent space can be implemented using an autoencoder architecture, which adaptively learns a brand-independent general identity feature representation by minimizing the joint loss function of reconstruction error and cross-domain distribution divergence.

[0037] As an optional embodiment, the logic for generating the adaptive privacy perturbation mask includes: Based on the security level of interactive events, basic permission gradients and high-order permission gradients are divided. Low-variance directional noise injection is configured for high-order permission gradients, and structured perturbation masks are configured for basic permission gradients. The noise injection intensity is positively correlated with the device heterogeneity in a nonlinear manner, which allows the hierarchical desensitization gradient vector to retain the high-frequency behavioral evolution information required for cross-domain verification while satisfying the hierarchical differential privacy budget constraint.

[0038] Based on the above, the logic for generating the adaptive privacy perturbation mask can be implemented using the terminal's built-in security coprocessor.

[0039] The system first divides the local update gradient into basic permission gradient and high-level permission gradient based on the security level of the interaction event.

[0040] For high-order permission gradients involving core account switching or payment verification, a low-variance directional noise injection module is configured to ensure that gradient signals of key identity dimensions are not completely submerged; while for basic interaction gradients, a structured perturbation mask is used for wide-area masking, thereby achieving differentiated grading of privacy protection strength at the gradient level.

[0041] Furthermore, the noise injection intensity exhibits a non-linear positive correlation with device heterogeneity. This means that the greater the brand difference between devices and the less uniform the underlying sensor architecture, the more the injected perturbation tends to retain high-frequency behavioral evolution information rather than static biological characteristics. A dynamic balancing strategy ensures that the hierarchical desensitization gradient vector can still support effective updates of the cross-domain validation model while strictly satisfying the hierarchical differential privacy budget constraint, thus avoiding model convergence stagnation caused by traditional fixed noise strategies.

[0042] In another alternative embodiment, the structured perturbation mask can be generated by a Laplace distribution or a Gaussian mixture model, and the perturbation bandwidth can be dynamically adjusted according to the real-time channel quality to adapt to the federated synchronization requirements in different network environments.

[0043] As an optional implementation, the calculation of dynamic aggregation weights includes constructing a joint cost function of topological trust decay function, data distribution offset metric and device brand heterogeneity penalty coefficient, and optimizing the joint cost function to determine the optimal contribution weight of each terminal node. When the local data distribution offset of any terminal node exceeds a preset deviation threshold, the optimal contribution weight is subjected to gradient pruning and channel isolation to suppress the parameter pollution of the global authentication model by heterogeneous distribution.

[0044] Based on the above, when calculating dynamic aggregation weights, the federated coordination node uses the optimizer to solve a joint cost function in real time, which includes the topology trust decay function, data distribution offset metric, and device brand heterogeneity penalty coefficient. When the local data distribution offset of any terminal node is detected to exceed a preset deviation threshold, the optimizer automatically triggers gradient pruning and channel isolation mechanisms, significantly reducing the aggregation weight corresponding to that node, or even removing it from the fusion channel of the current iteration cycle. Through the above dynamic suppression method, it is possible to effectively prevent abnormal data distribution (such as malicious tampering or extreme environmental noise) from polluting the parameters of the global authentication model.

[0045] Furthermore, the topological trust decay function performs exponential decay mapping based on the success rate of historical federated interactions, while the device brand heterogeneity penalty coefficient is obtained by calculating the cosine distance between the node's feature vector and the global benchmark. The two work together on the joint cost function to ensure that the allocation of high contribution weights is always biased towards terminal nodes with high credibility and data distribution close to the mainstream ecosystem.

[0046] As a result, the convergence stability of the global model in complex multi-terminal environments is substantially improved, avoiding the risk of model collapse caused by "bad money driving out good money" in traditional average aggregation strategies.

[0047] As an optional implementation, the logic for cross-domain feature deduction and dynamic takeover condition determination includes: Perform local forward propagation computation on the candidate takeover terminal side to generate a multidimensional confidence distribution vector; Real-time analysis of the current environment's risk entropy value and user behavior time-series inertial window; dynamically adjust the takeover confidence threshold range based on the risk entropy value. When the peak percentage of the multidimensional confidence distribution vector crosses the dynamically adjusted takeover confidence threshold and the entropy fluctuation range is within the preset stable range, the cross-domain feature deduction is deemed successful.

[0048] Based on the above, after receiving the global authentication model, the candidate takeover terminal will perform forward propagation calculation locally and output a multi-dimensional confidence distribution vector. To adapt to the identity takeover requirements in complex scenarios, the system will analyze the risk entropy value of the current environment in real time and construct a temporal inertial window based on the user's recent interaction behavior. Based on the risk entropy value, the takeover confidence threshold range is dynamically adjusted, so that the verification boundary is appropriately relaxed in low-risk environments (such as inside a home or in an authenticated office area), and the threshold is automatically tightened in high-risk environments (such as public places or unknown network access).

[0049] When the peak percentage of the multidimensional confidence distribution vector successfully crosses the dynamically adjusted takeover confidence threshold, and its entropy fluctuation range remains stable within a preset stable range, the system determines that the cross-domain feature deduction has passed. This mechanism avoids false blocking or excessive clearance caused by a single static threshold in visitor or dynamic scenarios by coupling the confidence distribution pattern with the environmental risk entropy.

[0050] In one alternative embodiment, the risk entropy value can be obtained by fusing ambient light signals, acoustic background, and near-field Bluetooth beacons using Kalman filtering, thereby providing a timely and context-aware input for threshold adjustment.

[0051] As an optional embodiment, the determination logic for the spatiotemporal continuity parameter includes: Construct the spatial proximity decay trajectory between the source terminal node and the candidate takeover terminal, and align the decay trajectory with the user behavior temporal inertial window through convolution; Calculate the rate of change of the derivative of the spatial proximity decay trajectory and the temporal overlap of the behavior characterized by convolution alignment; When the rate of change of the derivative is less than the preset smooth limit and the overlap of the behavior timing crosses the preset continuous limit, a seamless takeover instruction is generated.

[0052] Based on this, the determination logic of spatiotemporal continuity parameters mainly relies on the mathematical modeling of spatial proximity decay trajectories.

[0053] Furthermore, the system continuously collects the relative position signals of the source terminal node and the candidate takeover terminal, constructs a spatial proximity decay trajectory, and performs convolutional alignment processing on it with the user behavior temporal inertial window; by calculating the rate of change of the derivative of the decay trajectory and the behavioral temporal overlap represented by the convolutional alignment, the system can accurately characterize the continuity of the user's movement process.

[0054] Furthermore, when the rate of change of the derivative is less than a preset smooth limit and the overlap of the behavior sequence crosses a preset continuous limit, the system determines that the user has a clear intention to perform cross-terminal operations and then generates a seamless takeover command. This determination method based on the dual matching of kinematic features and behavioral patterns effectively filters out false trigger signals when the user accidentally approaches the backup device.

[0055] Furthermore, the convolution alignment process can employ a sliding time window mechanism, retaining only behavioral feature fragments that highly overlap with the current takeover window. This significantly reduces the state computation overhead caused by device switching, enabling session state snapshot transfer to be completed within millisecond latency.

[0056] As an optional embodiment, refer to the appendix Figure 2 A privacy-preserving federated learning update method for the smart terminal ecosystem also includes a non-registered entity intervention processing step: S5. When a secondary biometric subject not associated with the primary account is identified in the multimodal feature sequence, the static physiological components and short-term interaction patterns of the secondary biometric subject are extracted to generate a temporary shadow identity vector. S6. Project the temporary shadow identity vector to the preset low-dimensional permission subspace and bind it to an independent session lifecycle and instruction whitelist. During the lifecycle, only respond to the local gradient update request corresponding to the basic control instruction.

[0057] Furthermore, when a secondary biometric subject not associated with the primary account (such as a temporary visitor or a family member) is identified in the multimodal feature sequence, the system activates the shadow permission mapping mechanism.

[0058] Specifically, a lightweight feature extraction network captures the static physiological components and short-term interaction patterns of the secondary subject to generate a temporary shadow identity vector. The temporary shadow identity vector is then forcibly projected into a preset low-dimensional permission subspace and bound to an independent session lifecycle and instruction whitelist.

[0059] Furthermore, during its lifecycle, the system only allows the temporary shadow identity vector to respond to local gradient update requests corresponding to basic control commands (such as adjusting volume, querying the weather, and switching media sources), while higher-order commands involving core account configuration or access to sensitive data will be automatically intercepted by the underlying routing policy.

[0060] It is worth noting that the low-dimensional permission subspace and the main authentication model are physically isolated on the gradient aggregation channel to ensure that the interaction data of the secondary subject will not contaminate the global identity benchmark. When the session lifecycle expires or the main account's biometrics regain control of the interaction sequence, the temporary shadow identity vector automatically becomes invalid, and its associated gradient cache is cleared. This ensures a seamless interaction experience while achieving strict permission boundary control.

[0061] As an optional embodiment, refer to the appendix Figure 3 A privacy-preserving federated learning update system for the smart terminal ecosystem includes: The heterogeneous protocol bridging unit is configured to capture multimodal data streams from cross-brand terminal nodes, perform device protocol bias stripping and cross-domain feature space alignment, and output protocol-independent local identity representation vectors. The hierarchical gradient decoupling engine establishes a hardware-level isolated communication link with the heterogeneous protocol bridging unit. It has a built-in adaptive perturbation mask generator and directional noise injector to perform component separation and privacy masking of the local update gradient based on the security level of the interaction event and the heterogeneity of the device, and outputs a hierarchical desensitized gradient summary. The federated dynamic aggregation gateway integrates a cross-domain trust routing table and a joint cost optimizer. The cross-domain trust routing table maintains the trust decay trajectory and brand heterogeneous mapping relationship of each node. The joint cost optimizer reconstructs the aggregation channel topology based on the local data distribution offset and heterogeneous penalty coefficient, performs multi-source gradient fusion, and broadcasts the global authentication kernel. The scenario-adaptive takeover controller connects to the federated dynamic aggregation gateway and integrates the risk entropy assessment module and the spatiotemporal continuity matcher. When the risk entropy assessment module outputs a low-disturbance state and the spatiotemporal continuity matcher determines that the behavior sequence is coherent, it generates cross-node state transition control signaling and synchronizes the encrypted session context to the candidate takeover terminal.

[0062] Based on the above, in the above system architecture, the heterogeneous protocol bridging unit is usually deployed in the application processor layer or independent baseband chip of each smart terminal, and is configured to capture multimodal data streams of cross-brand terminal nodes; through the built-in modality alignment algorithm, the heterogeneous protocol bridging unit performs device protocol bias stripping and cross-domain feature space alignment, and directly outputs protocol-independent local identity representation vectors.

[0063] Furthermore, a hardware-level isolated communication link is established between the hierarchical gradient decoupling engine and the heterogeneous protocol bridging unit. The engine integrates an adaptive perturbation mask generator and a directional noise injector. When multimodal features flow through the engine, the system performs component separation and privacy masking on the local update gradient based on the security level of the interaction event and the heterogeneity of the device, and only outputs an encrypted hierarchical desensitized gradient summary.

[0064] Furthermore, the federated dynamic aggregation gateway, as the core hub in the cloud or at the edge, integrates a cross-domain trust routing table and a joint cost optimizer. The routing table maintains the trust decay trajectory and brand heterogeneous mapping relationship of each node in the form of a graph database, while the joint cost optimizer dynamically reconstructs the aggregation channel topology based on the real-time reported local data distribution offset and heterogeneous penalty coefficient. Under this architecture, the gateway performs multi-source gradient fusion and then broadcasts the global authentication kernel.

[0065] The context-adaptive takeover controller, connected to the federated dynamic aggregation gateway, continuously monitors environmental threats and user movement inertia around the terminal by integrating a risk entropy assessment module and a spatiotemporal continuity matcher. When the risk entropy assessment module outputs a low-disturbance state and the spatiotemporal continuity matcher determines that the behavior sequence is coherent, the controller immediately generates cross-node state transition control signaling and synchronizes the encrypted data stream containing the desensitized session context to the candidate takeover terminal, thereby achieving seamless transfer of cross-device permissions.

[0066] As an optional embodiment, the hierarchical gradient decoupling engine is deployed in the secure execution environment of the terminal device. The secure execution environment is configured with a gradient decoupling bus and a raw feature fuse to block the direct pass-through path of the un-de-sensitized raw multimodal sequence to the external network. Only the gradient summary after component separation and directional noise injection is allowed to be routed to the federated dynamic aggregation gateway through the gradient decoupling bus.

[0067] Optionally, the hierarchical gradient decoupling engine is deployed in the secure execution environment (TEE) or secure isolation zone (TrustZone) of the terminal device. This environment is physically isolated from the main operating system memory space in terms of hardware architecture. The secure execution environment is equipped with a dedicated gradient decoupling bus and a raw feature fuse. The fuse is connected in series between the sensor data output bus and the external communication interface for real-time monitoring of data flow.

[0068] Furthermore, when an un-anonymized raw multimodal sequence is detected attempting to initiate a transmission request to an external network, the circuit breaker immediately cuts off the path, thereby completely blocking the direct transmission of sensitive biometric data. Under this architectural constraint, only gradient summaries processed by component separation and directional noise injection are allowed to be routed to the federated dynamic aggregation gateway via the gradient decoupling bus.

[0069] As a result, the system's privacy protection mechanism has been moved from the software layer to the hardware layer. Even if the terminal's main operating system is compromised by malicious code, attackers will not be able to intercept the original identity data, which greatly improves the underlying data security of the smart terminal ecosystem during the federated update process.

[0070] As an optional embodiment, the scenario-adaptive takeover controller also includes a lightweight state packer and a near-field synchronization protocol stack, configured to compress the session context of the source terminal into an incremental state snapshot after cross-node state flow control signaling is triggered. The incremental state snapshot includes desensitized gradient routing mask parameters and dynamic takeover threshold identifiers, and is asynchronously broadcast to candidate takeover terminals via a near-field direct connection protocol or a local area network multicast channel.

[0071] Based on the above, the situation-adaptive takeover controller further integrates a lightweight state packer and a near-field synchronization protocol stack.

[0072] Furthermore, after the cross-node state transition control signaling is triggered, the state packer immediately extracts the current session context of the source terminal and performs serialization compression to generate an incremental state snapshot. The snapshot not only contains temporary interaction data of the application layer, but also strictly encapsulates the desensitized gradient routing mask parameters and dynamic takeover threshold identifier to ensure that the candidate takeover terminal can immediately inherit the verification context and privacy configuration parameters of the source terminal after receiving it.

[0073] Subsequently, the near-field synchronization protocol stack uses near-field direct connection protocols (such as StarFlash, Bluetooth Low Energy) or local area network multicast channels to send incremental state snapshots to candidate takeover terminals in an asynchronous broadcast manner. Since the snapshots only carry state increments rather than full model data, combined with the high bandwidth and low latency characteristics of near-field protocols, the synchronization overhead of cross-terminal session migration is minimized.

[0074] In an optional embodiment, if the candidate takeover terminal reports a failure to receive the data, the near-field synchronization protocol stack can automatically switch to the wide area network encrypted relay channel for retrying, ensuring the eventual consistent delivery of state transition instructions under complex network topology.

[0075] It should be noted that the above modules can be implemented by software or hardware. For the latter, they can be implemented in the following ways, but are not limited to: all the above modules are located in the same processor; or, the above modules are located in different processors in any combination.

[0076] According to another aspect of the present invention, an electronic device is also provided, the electronic device including a memory and a processor; the memory is used to store a program; the processor executes the program to implement the method of any of the foregoing.

[0077] Electronic devices can be terminals with edge computing capabilities, such as smartphones, tablets, smart wearables, or smart home hubs. Memory includes, but is not limited to, read-only memory (ROM), random access memory (RAM), flash memory, or solid-state drives, used to persistently store program instructions and intermediate state data for executing the above methods. The processor can be a general-purpose microprocessor, digital signal processor (DSP), application-specific integrated circuit (ASIC), or field-programmable gate array (FPGA), and it establishes a data interaction link with the memory through an internal system bus. When the processor loads and executes the program code in the memory, it invokes algorithm modules such as heterogeneous protocol decoupling, hierarchical gradient decoupling, and context-adaptive takeover to implement any of the aforementioned technical processes. Thus, this method can be instantiated and run on physical hardware, achieving local privacy protection and cross-domain collaborative updates without relying on a single cloud computing power.

[0078] According to another aspect of the present invention, a computer-readable storage medium is also provided, the storage medium storing a computer program that, when executed by a processor, implements the method of any of the foregoing.

[0079] Computer-readable storage media can be any non-transitory tangible carrier containing computer program code, such as a USB flash drive, external hard drive, optical disc, or embedded storage chip. When the computer program stored in the medium is read and executed by the processor of a terminal or server, it is transformed into a sequence of machine-executable instructions, thereby driving the hardware to complete core steps such as multimodal feature sequence acquisition, cross-domain feature space alignment, hierarchical privacy perturbation mask generation, and dynamic takeover threshold adjustment. By encapsulating this technical solution in an independent storage medium, the privacy-preserving federated learning update logic for the smart terminal ecosystem can be distributed, upgraded, and deployed without being tied to specific hardware, significantly improving the engineering portability and ecosystem compatibility of the technical solution.

[0080] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the method described in any of the foregoing.

[0081] Computer program products can take the form of software packages distributed in the cloud, system components built into an operating system, or functional plugins downloaded from an app store. When a computer program product is installed on a compatible electronic device, its program modules are invoked and executed by the processor in the system background or foreground, thus completely replicating the method steps of any of the aforementioned methods.

[0082] Furthermore, the product can continuously receive global authentication kernel updates from the federal coordination gateway via over-the-air (OTA) technology, and locally adaptively reconstruct heterogeneous protocol bridging and gradient decoupling logic, thereby ensuring that the terminal device maintains optimal cross-brand authentication performance and privacy security throughout its long lifecycle.

[0083] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A privacy protection federated learning updating method for an intelligent terminal ecosystem, characterized in that, Includes the following steps: The multimodal feature sequence and cross-domain device topology state parameters collected by the source terminal node are obtained. The multimodal feature sequence is decoupled from heterogeneous protocols and aligned with the cross-domain feature space to generate a protocol-independent local identity representation vector. Calculate the feature residual between the local identity representation vector and the historical global identity benchmark model, construct a local gradient update vector based on the feature residual, and generate an adaptive privacy perturbation mask according to the device heterogeneity and the security level of the interaction event. Perform targeted masking on the local gradient update vector to obtain a hierarchical desensitized gradient vector. The hierarchical desensitization gradient vector is routed to the federated coordination node. The federated coordination node calculates the dynamic aggregation weight based on the cross-domain trust decay factor, local data distribution offset and device brand heterogeneity penalty coefficient, performs gradient fusion operation, and outputs the global identity verification model for the next iteration cycle. The updated global authentication model is distributed to the candidate takeover terminal, and cross-domain feature inference is performed to generate a confidence distribution. When the confidence distribution meets the dynamic takeover conditions and the spatiotemporal continuity parameter between the source terminal and the candidate takeover terminal crosses the preset migration boundary, the session state snapshot transfer is triggered.

2. The privacy-preserving federated learning updating method for an intelligent terminal ecosystem according to claim 1, characterized in that, The steps of heterogeneous protocol decoupling and cross-domain feature space alignment include: Extract the device protocol bias component and the general biometric component from the multimodal feature sequence to construct a heterogeneous feature decoupling matrix; The heterogeneous feature decoupling matrix is ​​mapped to a cross-domain shared latent space by regularized orthogonal projection. The local identity representation vector is then iteratively updated using the feature cluster centers within the cross-domain shared latent space to remove the representation bias caused by the sampling differences of sensors from different brands of terminals.

3. The privacy-preserving federated learning updating method for an intelligent terminal ecosystem according to claim 1, characterized in that, The logic for generating the adaptive privacy perturbation mask includes: Based on the security level of the interactive event, a basic permission gradient and a high-order permission gradient are divided. Low-variance directional noise injection is configured for the high-order permission gradient, and a structured perturbation mask is configured for the basic permission gradient. The noise injection intensity is nonlinearly positively correlated with the device heterogeneity, so that the hierarchical desensitization gradient vector retains the high-frequency behavioral evolution information required for cross-domain verification while satisfying the hierarchical differential privacy budget constraint.

4. The privacy-preserving federated learning updating method for an intelligent terminal ecosystem according to claim 1, characterized in that: The calculation of the dynamic aggregation weight includes constructing a joint cost function of topological trust decay function, data distribution offset metric and device brand heterogeneity penalty coefficient, and optimizing the joint cost function to determine the optimal contribution weight of each terminal node. When the local data distribution offset of any terminal node exceeds a preset deviation threshold, the optimal contribution weight is subjected to gradient pruning and channel isolation to suppress the parameter pollution of the global authentication model by heterogeneous distribution.

5. The privacy-preserving federated learning updating method for an intelligent terminal ecosystem according to claim 1, characterized in that, The cross-domain feature deduction and dynamic takeover condition determination include: Perform local forward propagation computation on the candidate takeover terminal side to generate a multidimensional confidence distribution vector; The risk entropy value of the current environment and the time-series inertial window of user behavior are analyzed in real time, and the takeover confidence threshold range is dynamically adjusted based on the risk entropy value. When the peak percentage of the multidimensional confidence distribution vector crosses the dynamically adjusted takeover confidence threshold and the entropy fluctuation range is within the preset stable range, the cross-domain feature deduction is determined to be successful.

6. The privacy-preserving federated learning updating method for an intelligent terminal ecosystem according to claim 1, characterized in that, The determination logic for the spatiotemporal continuity parameter includes: Construct a spatial proximity decay trajectory between the source terminal node and the candidate takeover terminal, and align the decay trajectory with the user behavior temporal inertial window through convolution; Calculate the rate of change of the derivative of the spatial proximity decay trajectory and the temporal overlap of the behavior characterized by convolution alignment; When the rate of change of the derivative is less than a preset smoothing limit and the temporal overlap of the behavior crosses a preset continuous limit, a seamless takeover instruction is generated.

7. The privacy-preserving federated learning updating method for an intelligent terminal ecosystem according to claim 1, characterized in that, This also includes steps for handling cases involving non-registered entities: When a secondary biometric subject not associated with the primary account is identified in the multimodal feature sequence, the static physiological components and short-term interaction patterns of the secondary biometric subject are extracted to generate a temporary shadow identity vector. The temporary shadow identity vector is projected onto a preset low-dimensional permission subspace and bound to an independent session lifecycle and instruction whitelist. Within the lifecycle, it only responds to local gradient update requests corresponding to basic control instructions.

8. A privacy-preserving federated learning updating system oriented to an intelligent terminal ecosystem, applying the method according to any one of claims 1 to 7, characterized in that, include: The heterogeneous protocol bridging unit is configured to capture multimodal data streams from cross-brand terminal nodes, perform device protocol bias stripping and cross-domain feature space alignment, and output protocol-independent local identity representation vectors. The hierarchical gradient decoupling engine establishes a hardware-level isolated communication link with the heterogeneous protocol bridging unit. It has a built-in adaptive perturbation mask generator and a directional noise injector, which are used to perform component separation and privacy masking on the local update gradient according to the security level of the interaction event and the heterogeneity of the device, and output a hierarchical desensitized gradient summary. The federated dynamic aggregation gateway integrates a cross-domain trust routing table and a joint cost optimizer. The cross-domain trust routing table maintains the trust decay trajectory and brand heterogeneity mapping relationship of each node. The joint cost optimizer reconstructs the aggregation channel topology based on the local data distribution offset and heterogeneity penalty coefficient, performs multi-source gradient fusion, and broadcasts the global authentication kernel. The scenario-adaptive takeover controller, connected to the federated dynamic aggregation gateway, integrates a risk entropy assessment module and a spatiotemporal continuity matcher. When the risk entropy assessment module outputs a low-disturbance state and the spatiotemporal continuity matcher determines that the behavior sequence is coherent, it generates cross-node state transition control signaling and synchronizes the encrypted session context to the candidate takeover terminal.

9. The privacy-preserving federated learning updating system for the intelligent terminal ecosystem according to claim 8, characterized in that: The hierarchical gradient decoupling engine is deployed in the secure execution environment of the terminal device. The secure execution environment is configured with a gradient decoupling bus and a raw feature fuse to block the direct transmission path of the un-de-sensitized raw multimodal sequence to the external network. Only the gradient summary after component separation and directional noise injection is allowed to be routed to the federated dynamic aggregation gateway through the gradient decoupling bus.

10. The privacy-preserving federated learning updating system for an intelligent terminal ecosystem according to claim 8, wherein: The scenario-adaptive takeover controller also includes a lightweight state packer and a near-field synchronization protocol stack, configured to compress the session context of the source terminal into an incremental state snapshot after the cross-node state flow control signaling is triggered. The incremental state snapshot includes desensitized gradient routing mask parameters and a dynamic takeover threshold identifier, and is asynchronously broadcast to candidate takeover terminals through a near-field direct connection protocol or a local area network multicast channel.

Images