Active Membership Inference Test (aMINT) is a method for detecting whether specific data was used to train a machine learning model. aMINT proposes a novel multi-task learning process that simultaneously trains the original model (the Audited Model) and an auxiliary model (the MINT Model) that identifies training data. This approach is designed to integrate model auditability as an optimization goal during neural network training. The MINT layer uses intermediate activation maps as input and is trained to improve training data detection. When evaluating various neural networks, from MobileNet to Vision Transformer, on five public benchmarks, aMINT achieved over 80% accuracy in detecting data usage, significantly outperforming existing methods. aMINT and related methodological advancements contribute to increasing the transparency of AI models and establishing robust safeguards for security, privacy, and copyright protection.
