<ul><li>Traditional machine learning (ML) raises privacy concerns, while federated learning (FL) aims to protect data. However, FL training processes can still leak sensitive information, leading to privacy risks.</li><li>A two-stage defense method called AugMixCloak has been introduced to address membership inference attacks (MIA) in FL. This defense combines data augmentation and PCA-based information fusion, along with perceptual hashing, to protect query images.</li><li>Experimental results indicate that AugMixCloak effectively defends against both binary classifier-based MIA and metric-based MIA across multiple datasets and decentralized FL topologies.</li><li>Compared to other defense mechanisms like regularization-based approaches, AugMixCloak demonstrates stronger protection and better generalization abilities compared to confidence score masking.</li></ul>

AugMixCloak: A Defense against Membership Inference Attacks via Image Transformation

Discover more