RECALL: Towards Generalized Representations in Unsupervised Federated Learning Under Non-IID Conditions

This study investigates the impact of Non-IID data in Federated Learning (FL), focusing on unsupervised tasks related to reconstruction, such as image restoration, image denoising, and anomaly detection. Contrary to supervised learning, where performance is heavily influenced by data label distribution, we find that unsupervised learning in FL is more significantly affected by inter-class feature diversity. Our analysis reveals that datasets with low inter-class feature diversity, wherein representations of certain classes can be generalized to others, remain largely unaffected by Non-IID settings. Motivated by this insight, we propose the Representation-level CALibration aLgorithm (RECALL), a novel approach designed to calibrate latent representations in unsupervised models towards more generalized learning. RECALL leverages insights from previous global models to enhance local model training. Our experiments demonstrate the efficacy of RECALL in improving unsupervised learning tasks within the FedAVG framework under Non-IID conditions, underscoring the predominance of inter-class diversity over label distribution in unsupervised FL.