Federated Learning with Motion-Only Encryption for Visual Data Transformation in Human Activity Recognition

The growing deployment of intelligent video analytics in smart surveillance, healthcare, and autonomous systems has amplified concerns surrounding visual privacy. Traditional centralised learning requires transmitting raw video streams to cloud servers, creating significant privacy and security risks. Federated Learning (FL) offers a compelling alternative by enabling model training directly on local devices without exposing sensitive data. This work investigates the feasibility of privacy-preserving human activity recognition (HAR) in videos through a federated framework built on a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM)architecture. To ensure privacy by design, all experiments are conducted on a dataset in which Regions of Interest (ROIs) are encrypted, preventing the model from accessing identifiable visual content. We evaluate three FL aggregation strategies: FedAvg, FedAdam and FedProx, under IID (independent and identically distributed) and non-IID client partitions. The results demonstrate that FL-HAR achieves competitive performance, attaining up to 92% accuracy in challenging non-IID settings, with FedProx exhibiting superior stability under data heterogeneity. We further assess the robustness of the proposed system against inference-time adversarial manipulation using the Fast Gradient Sign Method (FGSM) attack. Quantitative and visual analysis show that ROI encryption substantially restricts the effectiveness of gradient-based perturbations, confining perturbations to background regions and preserving model confidence even at higher attack budgets. Overall, our study demonstrates that combining FL with ROI-level encryption provides an effective pathway toward secure, privacy-preserving, and accurate video-based HAR, making it suitable for deployment in ethically sensitive and security-critical environments.