Addressing Class Imbalance in Federated Learning

Federated learning (FL) is a promising approach for training decentralized data located on local client devices while improving efficiency and privacy. However, the distribution quantity of clients' side may lead to significant challenges such as class imbalance non-IID (non-independent identically distributed) data, which could greatly impact performance common model. While much effort has been devoted helping FL models converge when encountering issue not sufficiently addressed. In particular, executed by exchanging gradients in an encrypted form, completely observable either clients or server, previous methods do perform well FL. Therefore, it crucial design new detecting mitigating its impact. this work, we propose monitoring scheme that can infer composition each round, loss function -- Ratio Loss mitigate imbalance. Our experiments demonstrate importance acknowledging taking measures early possible training, effectiveness our method shown significantly outperform methods, maintaining