Privacy-Preserving and Cross-Domain Human Sensing by Federated Domain Adaptation with Semantic Knowledge Correction

Abstract

Federated Learning (FL) enables distributed training of human sensing models in a privacy-preserving manner. While promising, federated global models suffer from cross-domain accuracy degradation when the labeled source domains statistically differ from the unlabeled target domain. To tackle this problem, recent methods perform pairwise computation on the source and target domains to minimize the domain discrepancy by adversarial strategy. However, these methods are limited by the fact that pairwise source-target adversarial alignment alone only achieves domain-level alignment, which entails the alignment of domain-invariant as well as environment-dependent features. The misalignment of environment-dependent features may cause negative impact on the performance of the federated global model. In this paper, we introduce FDAS, a Federated adversarial Domain Adaptation with Semantic Knowledge Correction method. FDAS achieves concurrent alignment at both domain and semantic levels to improve the semantic quality of the aligned features, thereby reducing the misalignment of environment-dependent features. Moreover, we design a cross-domain semantic similarity metric and further devise feature selection and feature refinement mechanisms to enhance the two-level alignment. In addition, we propose a similarity-aware model fine-tuning strategy to further improve the target model performance. We evaluate the performance of FDAS extensively on four public and a real-world human sensing datasets. Extensive experiments demonstrate the superior effectiveness of FDAS and its potential in the real-world ubiquitous computing scenarios.