Time-Aware Security Intelligence for Federated Financial Systems: Deep Reinforcement Learning Against Temporal Poisoning Attacks

Financial institutions operating distributed machine learning systems face an emerging class of stealth adversaries who exploit temporal patterns across training cycles to inject persistent backdoors that remain dormant for months before activation. Unlike conventional single-round attacks, these sophisticated temporal poisoning strategies leverage sequential dependencies to bypass existing detection mechanisms while gradually compromising model integrity. Current defense frameworks remain fundamentally inadequate against such multi-period adversarial choreography, particularly in high-stakes financial environments where minute perturbations can trigger systemic failures. Existing security frameworks largely focus on static threat models and fail to address sophisticated multi-period adversarial strategies that unfold over time in financial transaction streams. To address these challenges, we propose DEFEND, a comprehensive defense framework that integrates temporal behavior analysis, robust statistical aggregation, and multi-scale verification into a unified multi-layer architecture. Our framework formulates defense coordination as a Markov Decision Process and employs Proximal Policy Optimization for adaptive policy learning that dynamically balances security enforcement with model utility. We design three sophisticated temporal attack models to comprehensively evaluate our defense mechanism: fixed-period data poisoning, multi-period data poisoning, and model weight poisoning attacks. The multi-layer defense architecture combines geometric median-based robust aggregation with Dynamic Time Warping pattern matching and adaptive client participation control. Extensive experiments on CIFAR-10, FEMNIST, and MNIST datasets demonstrate that DEFEND achieves superior defense performance with success rates of 95.6% for ResNet-18 and 94.0% for MobileNet V2, while maintaining clean accuracy levels between 85-95% across various data heterogeneity levels and malicious client ratios. Our framework provides theoretical guarantees for Byzantine robustness and practical scalability for moderate-scale federated deployments, making it well-suited for real-world financial applications requiring both security and efficiency.