Lightweight adversarial membership inference attack with temporal evolution against federated learning

Membership Inference Attacks (MIAs) pose a significant privacy threat to Federated Learning (FL) by enabling adversaries to exploit behavioral discrepancies between original and perturbed samples. However, most existing adversarial MIAs overlook the attack efficiency. To address this challenge, we propose ADVMIA, a novel adversarial attack framework that redefines the optimization perspective in this domain. ADVMIA is the first method to leverage the temporal evolution characteristics of FL models, offering a deeper understanding of the internal dynamics of adversarial sample generation. Concurrently, it introduces a threshold selection strategy that eliminates the need to train additional inference models, thereby reducing the high computational overhead associated with iterative adversarial optimization. We evaluate ADVMIA on two real-world datasets and compare its performance against five state-of-the-art MIAs. Experimental results demonstrate that ADVMIA achieves superior efficiency and heightened attack effectiveness. Furthermore, its robustness in practical FL scenarios is systematically validated through evaluations against two common defense mechanisms.