Adaptive Knowledge Recomposition for Personalized Federated Learning via Discrete Wavelet Transform

Zihao Lu, Junli Wang, Changjun Jiang

Abstract: Heterogeneous data silos hinder the application of deep learning in the Internet of Things. As a dealing scheme, personalized federated learning (pFL) distributedly customizes multiple local models for these silos. Most pFL methods directly use a global model to assist the local model optimization ignoring the performance drop caused by irrelevant or misleading global-model knowledge. To address this, we propose an adaptive knowledge recomposition approach (FedAKR), which refines relevant and correctly leading knowledge from the global model and recomposes it into the local model to promote personalization. Specifically, FedAKR provides a discrete wavelet transform-based method to recompose different kinds of knowledge in the same representation space. Facilitated by this common space, we introduce an enriched local optimization objective to establish a causal relationship between the refined global-model knowledge and recomposed local-model knowledge. The relationship guides effective and efficient knowledge refinement, thereby promoting personalization. Besides, we provide the theoretical proof of convergence for our novel pFL approach. Extensive experiments demonstrate that FedAKR achieves interpretable improvements, higher performance over 12 state-of-the-art methods, and the potential to further integrate pretrained large models.