Abstract
Federated learning supports decentralized model training while preserving user privacy, yet shared gradients contain sensitive information that exposes systems to privacy attacks. Among these, data reconstruction attacks, where adversaries recover training samples from gradients, pose critical threats. However, existing methods are often limited to small batch sizes and rely on activation patterns that rarely occur naturally, reducing their practical applicability. To address these challenges, we propose a novel attack framework called Targeted Neuron Stimulation (TNT), which systematically generates adversarial perturbations during training to enforce activation patterns conducive to effective data reconstruction. TNT solves a constrained optimization problem to inject subtle, imperceptible noise into target training batches, enabling an honest-but-curious server to extract sensitive data from client gradients with high fidelity. Through extensive evaluations across five datasets, including facial recognition and medical diagnosis, TNT expands reconstructable batch sizes from 8 to realistic levels and accelerates attack processes by up to 10 times, achieving significantly higher reconstruction fidelity compared to existing methods. By exploring latent vulnerabilities in gradient-sharing mechanisms, this work highlights the privacy risks inherent to federated learning systems and calls for stronger defenses to safeguard sensitive user data.