Critical Neuron-Based Backdoor Defense for Federated Learning: A Combined Dropout and Local Outlier Factor Approach

Federated learning enables collaborative model training without data sharing but remains vulnerable to backdoor attacks. Malicious participants can embed hidden triggers that cause targeted misclassification while preserving clean data performance. Existing defenses suffer from incomplete detection rates and significant utility degradation. We propose a dual-layered defense combining client-side neuron analysis with server-side anomaly detection. Our approach first computes criticality scores for individual neurons by analyzing gradient magnitude differences between backdoor and clean datasets. Critical neurons receive higher dropout probabilities during training, disrupting backdoor pathways while preserving benign functionality. Simultaneously, the server applies Local Outlier Factor analysis to model update parameters, filtering anomalous contributions before aggregation. Experimental evaluation on Fashion-MNIST and CIFAR-10 demonstrates attack success rate reduction from 89.2% to 3.7% while maintaining 98.3% clean accuracy. Compared to state-of-the-art methods, our defense achieves 38–40% better attack suppression with minimal utility loss. Theoretical analysis establishes convergence guarantees and bounds neuron criticality through loss function properties.