SL-Defender: Decentralized Defense for Adversarial Robustness in Federated Learning

Adversarial poisoning attacks are growing threats to federated deep learning systems, as models can be poisoned from any of the compromised participants. Existing solutions attempt to detect and defend these attacks by assuming that there is an honest central parameter aggregator on which the detection and defense techniques can be applied. This assumption can give rise to a single point of failure (SPOF) and an easy target for attackers to compromise the whole federated learning system. They also lack a robust mechanism to isolate the malicious node so as to protect the rest of the federated system. We propose SL-Defender, a novel solution that overcomes these limitations by using a decentralized method to detect and defend against attacks based on Swarm Learning. Our solution uses the selected layers from each of the federated learning participants and projects their parameters into a lower dimensional subspace and performs clustering analysis in this subspace. The proposed methodology adds the ability to act on identified malicious participants and contain their impact on the federated learning process or remove them if needed. The proposed SL-Defender thus eliminates the need for honest central aggregator and distributes computation to peers and dynamically monitors them and defends them from the federating poisoned parameters without interrupting the learning process. With the support of experimental results on MNIST Dataset, we prove that SL-Defender will offer a unique position to improve the reliability and trustworthiness of Swarm Learning and Federated Machine learning in general.