FedLCD: A Federated Learning Framework with Layer-Wise Consistency Detection

Abstract

Federated learning enables multiple clients to collaboratively train a global model without sharing raw data, but it is vulnerable to Byzantine attacks, in which malicious clients upload manipulated model updates to disrupt training or degrade model accuracy. Existing Byzantine-robust aggregation methods typically operate at the whole-model level, making them less effective against partial-layer attacks and causing potentially useful information from unaffected layers to be discarded. To address this, we propose a Federated Learning framework with Layer-wise Consistency Detection (FedLCD). Specifically, we design a layer-wise Byzantine robust aggregation algorithm that detects malicious updates by computing pairwise cosine similarities at the granularity of individual layers, and retains only those with higher average similarity. This design filters malicious updates while preserving benign information from unaffected layers. Theoretical analysis confirms that FedLCD achieves a convergence rate of O(1/T), where T denotes the total number of communication rounds. This rate matches that of standard FedAvg, while FedLCD guarantees Byzantine robustness when benign clients constitute the majority. Extensive experiments on the CIFAR-10 dataset demonstrate that FedLCD achieves accuracy comparable to or higher than existing baselines across various Byzantine attack scenarios, while maintaining robust detection performance, particularly under partial-layer attacks.