DropBlock: A structured regularisation technique for mitigating membership inference attacks (MIAs) in machine learning models

This study proposes DropBlock, a structured regularization technique, to mitigate Membership Inference Attacks (MIAs). Unlike traditional dropout, DropBlock masks contiguous regions of feature maps during training, disrupting spatial correlations and preventing data memorization. Evaluated on MNIST, CIFAR-10, and CIFAR-100, DropBlock significantly outperforms baseline defenses, reducing MIA success rates by up to 12.8% on CIFAR-10 and 12.1% on CIFAR-100. Furthermore, it suppresses the attack True Positive Rate by over 40% at low false alarm rates while improving classification accuracy by 2–3%. These findings confirm that DropBlock provides a superior and practical defense without requiring noise injection. 2018 The Korean Institute of Communications and Information Sciences. Publishing Services by Elsevier B.V. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/by-nc-nd/4.0/ ).