TMN: An Efficient Robust Aggregator for Federated Learning

The collaboration of multiple organizations, such as hospitals, with access to data, can expedite the training process, resulting in superior machine learning models with increased data availability. However, the sensitivity of medical data poses challenges to information sharing without compromising privacy and confidentiality. Federated Learning (FL) offers a promising solution by enabling collaborative training through a data-sharing-free approach. Nevertheless, a large number of FL aggregation algorithms assume clients are honest, leaving the global model vulnerable to poisoning attacks. Approaches to safeguard against such attacks often add high computational costs, making them unsuitable for practical applications. In this work, we propose a robust aggregation rule, named Trimmed-Median Neighbourhood, for Byzantine-tolerant machine learning, offering computational efficiency and resilience to various attacks. Our method achieves up to a 2% improvement over the baseline and modified approaches in an adversarial attack setting on a non-IID data split from the HAM10000 dataset while maintaining low computational requirements. The code is available here.