Self-Correcting Client-Initiated Gradient Inversion via Adaptive Diffusion Control

Federated learning (FL) ostensibly safeguards data privacy by sharing only model updates. However, gradient inversion attacks (GIAs) have exposed critical vulnerabilities in this paradigm. While prior literature predominantly assumes a privileged server-side adversary, this work addresses the more insidious and practical threat of client-initiated attacks. Existing client-initiated methodologies remain fragile, treating inversion as a static, single-shot optimization problem that frequently succumbs to catastrophic optimization collapse due to the stochastic nature of FL dynamics. To overcome these limitations, we propose DCPT (i.e., Diffusion-driven Client-initiated Privacy Theft), a resilient framework that shifts the paradigm toward multi-stage adaptive inversion. Specifically, DCPT integrates a novel adaptive control framework with a closed-loop feedback mechanism, monitoring reconstruction fidelity to dynamically alternate between Exploitation and Exploration, thereby achieving self-correcting. To further guarantee robustness against gradient drift, we incorporate a rolling history constraint for temporal regularization and a bilevel optimization for on-the-fly prior adaptation. Extensive experiments on MNIST and CIFAR-10 demonstrate that DCPT achieves state-of-the-art reconstruction fidelity. Crucially, it also achieves unprecedented multi-round stability and resilience, highlighting the urgent need for advanced defenses against generative client-initiated adversaries.