Federated Unlearning for Human Activity Recognition

TL;DR

This paper tackles post-training privacy concerns in Federated HAR by introducing a lightweight federated unlearning method. It uses a KL-divergence–based fine-tuning objective that aligns the forgotten data distribution with a third-party dataset, while preserving performance on remaining data via a calibrated loss weight $\lambda$. A membership inference evaluation framework validates forgetting, and experiments across HAR70+, HARTH, and MNIST (iid and non-iid) demonstrate unlearning accuracy comparable to retraining with speedups from $294\times$ to $6119\times$. The study also analyzes third-party data selection (random noise vs retained client data) and shows generally robust forgetting, with some sensitivity under non-iid conditions. Overall, the work provides a practical, privacy-preserving unlearning mechanism for HAR in FL and highlights considerations for data-source choice and distributional heterogeneity.

Abstract

The rapid evolution of Internet of Things (IoT) technology has spurred the widespread adoption of Human Activity Recognition (HAR) in various daily life domains. Federated Learning (FL) is frequently utilized to build a global HAR model by aggregating user contributions without transmitting raw individual data. Despite substantial progress in user privacy protection with FL, challenges persist. Regulations like the General Data Protection Regulation (GDPR) empower users to request data removal, raising a new query in FL: How can a HAR client request data removal without compromising other clients' privacy? In response, we propose a lightweight machine unlearning method for refining the FL HAR model by selectively removing a portion of a client's training data. Our method employs a third-party dataset unrelated to model training. Using KL divergence as a loss function for fine-tuning, we aim to align the predicted probability distribution on forgotten data with the third-party dataset. Additionally, we introduce a membership inference evaluation method to assess unlearning effectiveness. Experimental results across diverse datasets show our method achieves unlearning accuracy comparable to \textit{retraining} methods, resulting in speedups ranging from hundreds to thousands.

Figures (9)

• Figure 1: Federated unlearning framework for HAR.
• Figure 2: Federated unlearning method.
• Figure 3: Membership inference evaluation method.
• Figure 4: Data distribution for each client across various datasets, with distinct colors denoting different labels.
• Figure 5: HAR70+