Exploiting Layerwise Feature Representation Similarity For Backdoor Defence in Federated Learning

Federated learning is an emerging paradigm for distributed machine learning that enables clients to collaboratively train models while maintaining data privacy. However, this approach introduces vulnerabilities, notably the risk of backdoor attacks where compromised models may perform normally on clean data but maliciously on poisoned inputs. A range of defences has been proposed in the literature based on robust aggregation, differential privacy or certified robustness and clustering/trust score-based approaches. In this work, we introduce FedAvgCKA, a novel defence mechanism that leverages the learned representations of neural networks to distinguish between benign and malicious submissions from clients. We demonstrate the effectiveness of FedAvgCKA across various federated learning scenarios and datasets, showcasing its ability to maintain high main task accuracy and significantly reduce backdoor attack success rates even in non-iid settings.