Fair detection of poisoning attacks in federated learning on non-i.i.d. data

Reconciling machine learning with individual privacy is one of the main motivations behind federated learning (FL), a decentralized machine learning technique that aggregates partial models trained by clients on their own private data to obtain a global deep learning model. Even if FL provides stronger privacy guarantees to the participating clients than centralized learning collecting the clients' data in a central server, FL is vulnerable to some attacks whereby malicious clients submit bad updates in order to prevent the model from converging or, more subtly, to introduce artificial bias in the classification (poisoning). Poisoning detection techniques compute statistics on the updates to identify malicious clients. A downside of anti-poisoning techniques is that they might lead to discriminate minority groups whose data are significantly and legitimately different from those of the majority of clients. This would not only be unfair, but would yield poorer models that would fail to capture the knowledge in the training data, especially when data are not independent and identically distributed (non-i.i.d.). In this work, we strive to strike a balance between fighting poisoning and accommodating diversity to help learning fairer and less discriminatory federated learning models. In this way, we forestall the exclusion of diverse clients while still ensuring detection of poisoning attacks. Empirical work on three data sets shows that employing our approach to tell legitimate from malicious updates produces models that are more accurate than those obtained with state-of-the-art poisoning detection techniques. Additionally, we explore the impact of our proposal on the performance of models on non-i.i.d local training data.