Privacy-preserving personalized human activity recognition via federated learning, adaptive differential privacy, and explainable AI

Human Activity Recognition (HAR) from wearable inertial sensor data constitutes a foundational technology for intelligent health monitoring, ambient assisted living, and pervasive computing. Despite remarkable accuracy advances enabled by deep neural architectures, conventional HAR systems predominantly rely on centralized training paradigms that aggregate raw personal mobility streams on remote servers, creating severe privacy vulnerabilities, regulatory hurdles, and clinical adoption barriers. To address these challenges, this paper presents FedXHAR, a privacy-preserving, personalized, and interpretable federated learning framework designed specifically for heterogeneous edge devices. FedXHAR coordinates three core technical components: a Shapley-guided attentive aggregation mechanism that weights client gradient contributions based on locally computed channel-importance summaries to mitigate client drift under non-IID distributions; an adaptive differential privacy gradient shield combining privately tracked quantile-based clipping with an inverse-root noise decay schedule and Rényi Differential Privacy accounting to balance privacy guarantees and model utility; and an on-device first-order attribution engine paired with a decoupled dual-stage personalization workflow (FedXHAR-L) that generates local feature saliency maps in a single backward pass while fine-tuning client-specific decision boundaries without sharing raw sensor data. Extensive empirical evaluations across four benchmark datasets—UCI HAR, PAMAP2, WISDM, and USC-HAD—demonstrate that FedXHAR achieves state-of-the-art recognition accuracies of 97.84%, 96.21%, 95.73%, and 94.90%, outperforming leading federated baselines by up to 5.29 percentage points while cutting convergence rounds by 47.6% and yielding a 2.73 $$\times$$ reduction in total communication volume. Under a strict differential privacy budget of $$\varepsilon=1.0$$ , FedXHAR surpasses standard private federated averaging by 10.49 percentage points on UCI HAR, while rigorous leave-one-subject-out evaluations confirm robust out-of-distribution generalization across unseen wearers. These theoretical and empirical results demonstrate that FedXHAR offers a reliable, communication-efficient, and privacy-compliant foundation for transparent human activity recognition at the edge.

Authors

Institutions

Publication Details

Journal
Discover Computing
Published
2026-11-18
DOI
https://doi.org/10.1007/s10791-026-10834-2
Primary Topic
Privacy-Preserving Technologies in Data
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Privacy-preserving personalized human activity recognition via federated learning, adaptive differential privacy, and explainable AI

Bholanath Roy, Aman Kumar Patel
2 links
Discover Computing
Privacy-Preserving Technologies in Data
article

Privacy-preserving personalized human activity recognition via federated learning, adaptive differential privacy, and explainable AI

Bholanath Roy, Aman Kumar Patel
article en

Abstract

Human Activity Recognition (HAR) from wearable inertial sensor data constitutes a foundational technology for intelligent health monitoring, ambient assisted living, and pervasive computing. Despite remarkable accuracy advances enabled by deep neural architectures, conventional HAR systems predominantly rely on centralized training paradigms that aggregate raw personal mobility streams on remote servers, creating severe privacy vulnerabilities, regulatory hurdles, and clinical adoption barriers. To address these challenges, this paper presents FedXHAR, a privacy-preserving, personalized, and interpretable federated learning framework designed specifically for heterogeneous edge devices. FedXHAR coordinates three core technical components: a Shapley-guided attentive aggregation mechanism that weights client gradient contributions based on locally computed channel-importance summaries to mitigate client drift under non-IID distributions; an adaptive differential privacy gradient shield combining privately tracked quantile-based clipping with an inverse-root noise decay schedule and Rényi Differential Privacy accounting to balance privacy guarantees and model utility; and an on-device first-order attribution engine paired with a decoupled dual-stage personalization workflow (FedXHAR-L) that generates local feature saliency maps in a single backward pass while fine-tuning client-specific decision boundaries without sharing raw sensor data. Extensive empirical evaluations across four benchmark datasets—UCI HAR, PAMAP2, WISDM, and USC-HAD—demonstrate that FedXHAR achieves state-of-the-art recognition accuracies of 97.84%, 96.21%, 95.73%, and 94.90%, outperforming leading federated baselines by up to 5.29 percentage points while cutting convergence rounds by 47.6% and yielding a 2.73 $$\times$$ reduction in total communication volume. Under a strict differential privacy budget of $$\varepsilon=1.0$$ , FedXHAR surpasses standard private federated averaging by 10.49 percentage points on UCI HAR, while rigorous leave-one-subject-out evaluations confirm robust out-of-distribution generalization across unseen wearers. These theoretical and empirical results demonstrate that FedXHAR offers a reliable, communication-efficient, and privacy-compliant foundation for transparent human activity recognition at the edge.

Discover ComputingVol. 29(1)
Maulana Azad National Institute of Technology (IN)
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon