DEFEND: Intelligent Temporal Backdoor Detection and Mitigation in Federated Learning via Reinforcement Learning-Coordinated Multi-Layer Defense

Collaborative machine learning in financial systems faces an escalating security threat: temporal backdoor attacks that exploit multi-round dependencies to systematically compromise fraud detection and risk assessment models—a challenge that existing static defense mechanisms cannot adequately counter. This paper presents DEFEND (DEep Federated Ensemble Network Defense), a comprehensive framework that integrates multi-layer defense with reinforcement learning-based adaptive coordination to counter sophisticated temporal backdoor strategies in federated learning environments. The framework introduces four key innovations: (1) Temporal Behavioral Analysis Layer employing multi-scale statistical profiling with dynamic time warping for attack pattern recognition across communication rounds, (2) Byzantine-Robust Statistical Aggregation using geometric median estimation with adaptive outlier detection, (3) Multi-Scale Validation Protocol with automated model rollback mechanisms, and (4) MDP-based Defense Coordination formulating security decisions as a Markov Decision Process optimized via Proximal Policy Optimization to dynamically balance robustness and utility. Extensive experiments on the FinMultiTime dataset across three distinct market periods (2009-2025) demonstrate superior performance over state-of-the-art baselines, achieving defense success rates of 95.6\%$\pm$1.0\% for ResNet-18 and 94.0\%$\pm$1.2\% for MobileNet-V2 while maintaining clean accuracy above 85\%. Ablation studies reveal that the MDP-based coordination provides the largest individual contribution (8.2\% defense success rate improvement), while the complete multi-layer architecture achieves up to 18.7\% improvement over single-layer baselines. Cross-period generalization analysis demonstrates robust transferability with less than 6\% performance degradation across different market regimes, validating practical deployment viability in dynamic financial environments.