Member Reasoning Attack Detection and Privacy Disclosure Quantification in Federated Learning Gradient Swap Process

Abstract

Federated learning aims to protect user data privacy by training locally and exchanging only model updates. However, the gradient of the exchange itself can reveal information that allows malicious actors to launch member-inference attacks to infer whether a particular data sample is present in a client’s training set, posing a serious privacy threat. In this paper, a complete detection and quantification scheme is proposed for the member inference attack in federated learning gradient exchange. First, we design an attack detection framework based on gradient feature analysis, which uses ensemble learning model to identify potential malicious attacks by extracting and aggregating high-dimensional statistical features and distribution anomalies of client-uploaded gradients. Secondly, in order to evaluate the severity of privacy leakage when the attack succeeds, we propose a privacy leakage quantification method. By measuring the sensitivity difference of gradient to member and non-member samples, we construct a computable leakage scoring system, and classify different leakage risk levels accordingly. Experimental results show that the detection method proposed in this paper can effectively identify member inference attacks in various federated learning scenarios, and the quantitative disclosure index can accurately reflect the privacy risk level under different attack configurations and model states. This study provides theoretical and technical references for privacy security assessment and dynamic protection of federated learning systems.