gMINT: Gradiant-based Membership Inference Test applied to Image Models.

Daniel DeAlcala, Aythami Morales, Julian Fierrez, Gonzalo Mancera, Ruben Tolosana; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 2781-2790

Membership Inference Test (MINT) is a method designed to determine whether specific data samples were used during model training. This paper explores the use of gradient-based information for MINT (gMINT). In particular, we leverage the gradient updates applied to each weight during the learning process, which we refer to as Weight Modifiers. We systematically analyze different types of Weight Modifiers and propose various approaches to process them effectively. Our experiments are conducted across multiple state-of-the-art architectures and four benchmark datasets, demonstrating that our method achieves near-perfect detection (100% in most scenarios). These results highlight the effectiveness of Weight Modifiers as a key signal for Membership Inference, further advancing the field of model auditing and privacy assessment in AI systems.