A federated learning framework termed FedXHAR achieves state-of-the-art recognition accuracies across benchmark wearable sensor datasets while enforcing formal client-level differential privacy guarantees and providing transparent on-device feature attributions. Experimental evaluations conducted on four standardized inertial measurement datasets—UCI HAR, PAMAP2, WISDM, and USC-HAD—demonstrate that the architecture attains classification accuracies of 97.84 percent, 96.21 percent, 95.73 percent, and 94.90 percent, respectively. These figures surpass leading federated learning baselines by up to 5.29 percentage points while slashing required communication convergence rounds by 47.6 percent and yielding an overall 2.73-fold reduction in total communication payload. Under an exacting differential privacy budget of epsilon equals 1.0, the framework outperforms standard differentially private federated averaging by 10.49 percentage points on the UCI HAR benchmark. Rigorous leave-one-subject-out cross-validations confirm that the system maintains robust out-of-distribution generalization when applied to completely unseen human bodies, demonstrating that distributed edge models can overcome user behavioral variance without harvesting raw personal mobility traces onto centralized cloud servers.
Automated inference of physical movements—including walking, running, sitting, ascending stairs, and domain-specific motor gestures—from miniaturized inertial measurement units embedded in smartphones and smartwatches has become foundational for continuous health monitoring, neurological disorder assessment, and ambient assisted living. Although modern centralized deep neural networks readily achieve benchmark recognition accuracies exceeding 95 percent by extracting hierarchical temporal representations from raw signals, centralized collection creates severe privacy vulnerabilities. Longitudinal kinematic streams can expose sensitive medical conditions, functional motor decline, and granular lifestyle habits, directly conflicting with privacy regulations such as the European General Data Protection Regulation and the Health Insurance Portability and Accountability Act. While standard federated learning circumvents raw data centralization by transmitting local model updates to a coordination server, real-world deployment on wearable devices encounters acute statistical heterogeneity. Non-identical activity profiles among users induce client drift, which pulls local optimization directions apart and destabilizes global convergence, while naive differential privacy perturbation severely degrades recognition utility.
To overcome these systemic vulnerabilities, the FedXHAR architecture operates across client and server nodes by coupling three coordinated technical modules: an on-device hybrid neural recognizer, an adaptive differential privacy gradient shield, and a server-side Shapley-guided attentive aggregation engine. Each client device processes sliding-window multivariate sensor segments of 128 timesteps with a 50 percent overlap through a hybrid convolutional and recurrent network totaling 847,328 trainable parameters. The feature extraction backbone integrates a dual-branch convolutional neural network containing a narrow-kernel path with filter length three to capture rapid motor transitions and a wide-kernel path with filter length seven to model cyclical kinetic patterns, each utilizing 64 filters alongside batch normalization. Two stacked bidirectional long short-term memory layers with a hidden dimension of 128 capture temporal dependencies across both forward and reverse directions, followed by a channel-wise self-attention mechanism that directs representational capacity toward discriminative temporal segments before feeding a fully connected classification head trained under cosine annealing.
A critical innovation within the framework is its Shapley-guided attentive aggregation mechanism, which discards uniform client averaging in favor of an importance-weighted fusion calculated from localized cooperative game theory. To establish aggregation weights without exposing raw sensor windows, each participant executes a client-side FastSHAP explainability pipeline over its local validation partition, generating an M-dimensional channel-importance summary vector whose elements represent the mean absolute attribution of each sensor channel. Because this explanatory summary constitutes a sensitive disclosure, the client bounds its sensitivity by clipping its Euclidean norm to a threshold of 1.0 before perturbing the vector with calibrated spherical Gaussian noise. At the central coordination server, attention weights are derived by computing the temperature-scaled cosine similarity between each privatized client profile and the global population mean under a temperature factor of 0.5. Theoretical convergence analysis proves that this weighting mechanism reduces expected gradient divergence under non-identical client distributions by steering global parameter updates toward a consensus directional alignment.
The framework stabilizes distributed optimization against noisy gradient perturbations through an adaptive differential privacy gradient shield. Rather than relying on constant clipping thresholds that artificially truncate large initial updates or drown small late-stage updates in excessive noise, the shield dynamically tracks gradient magnitudes using a private geometric quantile tracker. During each federation round, sampled clients compute their accumulated model update over five local training epochs and transmit a single noisy indicator bit revealing whether their unperturbed update fell below the broadcast threshold. The server updates the clipping boundary once per round toward an empirical unclipped target fraction of 0.85 within strictly bounded limits between 0.1 and 10.0. Simultaneously, the effective Gaussian noise multiplier follows an inverse-root decay schedule starting from a calibrated initial level and decaying across successive communication rounds, preventing premature parameter stalling while ensuring that noise injection diminishes as local client weights converge toward shared stationary regions.
Privacy accounting within the system is rigorously formulated under client-level differential privacy using add-or-remove adjacency, protecting the entirety of a user’s sensor windows as well as their presence or absence in training. Because local training proceeds via ordinary non-private optimization and perturbation occurs exactly once per federation round upon the accumulated pseudo-gradient and channel profile, privacy loss does not accumulate over internal mini-batch steps or training epochs. Cumulative privacy consumption is tracked using exact numerical evaluations of sampled-Gaussian Rényi Differential Privacy across integer orders ranging from 2 through 256, accounting for Poisson client participation subsampling at rate q equals m divided by K. The resulting Rényi divergence composes additively over communication rounds and over the two independent Gaussian transmissions per round—the joint clipped update with quantile bit and the channel-importance summary—before undergoing conversion into standard epsilon and delta bounds, ensuring verifiable formal protection against inference attacks observing the sequence of global models.
To guarantee operational viability over bandwidth-constrained mobile and wireless links, FedXHAR incorporates a biased Top-K gradient sparsification operator paired with on-device error feedback. Each client retains only the 20 percent largest coordinates of its privatized update vector, setting the remaining 80 percent of dimensions to zero while accumulating the residual quantization differences within an internal memory buffer to correct subsequent optimization steps. Transmitting only the top 169,466 parameters encoded alongside 32-bit coordinate indices, in combination with the low-dimensional privatized channel summary vector of 24 to 108 bytes, restricts the uplink payload to 1.356 megabytes per round compared to 3.389 megabytes for dense federated baselines. When factoring in the downlink transmission of the dense global model and scalar clipping threshold, total round-trip exchange requires 4.745 megabytes per client round. Combined with accelerated convergence that reaches target accuracy in 43 iterations instead of 82, the framework reduces total consumed communication volume from 2.14 gigabytes to 0.23 gigabytes on the UCI HAR benchmark.
The empirical performance of the framework was validated across three distinct data partitioning schemes evaluated over ten independent random seeds, utilizing paired t-tests and Wilcoxon signed-rank tests adjusted via the Holm–Bonferroni method. Under Protocol A, where each client strictly represents a unique physical participant without cross-subject window leakage, the system maintained recognition performance within 0.7 to 1.4 percentage points of an unconstrained centralized upper bound. Under Protocol B, which partitions pooled sensor windows into non-identical Dirichlet shards across populations scaling from 50 to 200 clients to stress-test aggregation under extreme label skew, FedXHAR experienced an accuracy drop of only 1.77 percentage points on UCI HAR, compared to severe degradations of 4.92 and 3.71 percentage points observed under standard federated averaging and pFedMe. Protocol C executed comprehensive leave-one-subject-out cross-validation, demonstrating that the global model exhibits superior zero-shot transfer across completely held-out individuals, confirming that importance-weighted aggregation prevents the global network from over-indexing on dominant user cohorts.
To address physiological disparities and distinct biomechanical gait characteristics, the framework implements a dual-stage personalization workflow designated as FedXHAR-L. During the initial collaborative phase, participant devices coordinate to optimize the shared convolutional and recurrent representation trunk. Upon global convergence, clients download the converged parameters, permanently freeze the feature extraction trunk, and execute three local fine-tuning epochs exclusively on their private classification head with a learning rate of 0.0005. Under subject-independent leave-one-out testing, this localized calibration elevates classification accuracy to 96.84 percent on UCI HAR and 94.71 percent on PAMAP2. Furthermore, per-class evaluations across all 18 movement categories of the PAMAP2 dataset confirm marked performance elevations for transitional and low-frequency activities such as rope jumping and ascending stairs, where Shapley-guided aggregation effectively consolidates minority-class representations from clients who execute those specific motor tasks more frequently.
A systematic component ablation conducted on the UCI HAR and PAMAP2 benchmarks isolates the statistical contributions of each architectural module under a uniform privacy budget of epsilon equals 10.0 and delta equals 10 to the negative fifth. Transitioning from fixed-threshold clipping to the adaptive quantile tracking shield and inverse-root noise decay restores utility by 3.31 percentage points on UCI HAR and 4.08 percentage points on PAMAP2. Subsequent enforcement of 20 percent Top-K gradient compression introduces an imperceptible accuracy variation of minus 0.36 percentage points while eliminating 80 percent of gradient payload. Incorporating Shapley-guided attentive aggregation provides the single largest performance leap, increasing accuracy by 6.07 and 9.05 percentage points while reducing convergence rounds from 73 to 43. Finally, activating on-device classification head calibration yields an incremental gain of up to 0.69 percentage points, illustrating how decoupled local tuning refines subject-level boundaries without altering shared representational parameters.
Hardware execution metrics were profiled directly on physical mobile platforms, including the Qualcomm Snapdragon 888 and MediaTek Dimensity 900 mobile processors alongside the NVIDIA Jetson Nano edge computing module. By formulating on-device explainability through a first-order Taylor approximation that requires only a single backward pass, attribution saliency maps achieve a faithfulness correlation score of 0.912 on PAMAP2 while executing in 43.7 milliseconds on the Snapdragon 888 processor, outperforming Integrated Gradients by 79.7 percent in latency without auxiliary memory allocations. Although per-round client training on Snapdragon hardware requires 5.73 seconds compared to 3.41 seconds for standard federated averaging due to backpropagation demands, FedXHAR reduces total cumulative device-active time by 11.9 percent because it achieves convergence in substantially fewer communication rounds. Peak client random-access memory consumption remains capped at 267 megabytes with a compact checkpoint footprint of 3.24 megabytes.
The authors delineate specific structural boundaries regarding the interpretation of their findings. The reported privacy accounting and classification accuracy trajectories were generated through simulated federated communication using the Flower framework on dual NVIDIA A100 graphics processing units, which faithfully models network latency and client dropouts but does not capture asynchronous network stragglers, true edge clock drift, or malicious adversarial poisoning attacks. Additionally, the first-order gradient attribution method provides a linear local approximation that may miss complex multivariate feature interactions within dense sensor arrays, while the privatized channel-importance vectors introduce modest estimation variance into server attention weights. Finally, the empirical hardware validation establishes operational feasibility for smartphone-class system-on-chip architectures and embedded Linux hardware, leaving ultra-low-power microcontrollers operating under sub-megabyte memory limits for future research involving integer quantization and model pruning.
Subject of Research: Privacy-preserving federated human activity recognition using adaptive differential privacy and explainable artificial intelligence
Article Title: Privacy-preserving personalized human activity recognition via federated learning, adaptive differential privacy, and explainable AI
Article References: Patel, A. K., & Roy, B. (2026). Privacy-preserving personalized human activity recognition via federated learning, adaptive differential privacy, and explainable AI. Discover Computing, 29(1), Article 905. https://doi.org/10.1007/s10791-026-10834-2
Image Credits: AI Generated
DOI: 10.1007/s10791-026-10834-2
Keywords: human activity recognition, federated learning, differential privacy, explainable AI, FastSHAP, wearable sensors, edge computing, inertial measurement units, client drift, gradient sparsification, Rényi differential privacy, quantile clipping, biometric privacy, deep learning, convolutional neural networks, bidirectional LSTM
News Source: Denise Maddox. (November 18, 2026). FedXHAR Framework Achieves Accurate Activity Recognition While Safeguarding Edge Wearer Privacy. Scienmag.



