Non-IID Data Based Federated Learning Using Mutual Knowledge Distillation and Generative Adversarial Network

Key words:  federated learning  non-independently and identically distributed  knowledge distillation  generative adversarial network

Abstract:

User heterogeneity in Federated Learning (FL) necessitates the re-optimization of local models, results in the loss of global knowledge, and leads to slow convergence and degraded performance. When dealing with heterogeneous clients in FL, Knowledge Distillation (KD) is a standard approach to increasing efficiency and improving generalization. However, KD relies on proxy datasets, and the underutilization of client knowledge in guiding local model learning has a negative impact on the quality of the aggregation model. Regarding these, a new FL method is proposed based on a Generative Adversarial Network (GAN) and KD, which has two training stages. At the first stage, client collaboration pre-trains a GAN to generate secondary datasets, overcoming the limitations of proxy datasets. In the second stage, the mutual KD process is implemented through the dynamic adjustment of the weights of client models to tackle the underutilization of integrated client knowledge. In the training phase, the pre-trained generator after fine-tuning can transfer knowledge from multiple local models to a global model to enhance the efficiency of KD. On the introduced benchmark datasets, the experimental results show that the proposed FL method needs fewer communication rounds and reflects better generalization than the state-of-the-art FL methods.