Design and Implementation of an MTD-based Mitigation Approach for Membership Inference Attacks

With the rapid growth of Internet of Things (IoT) devices, data privacy concerns have become increasingly prominent. Although federated learning can reduce privacy risks, attackers can still successfully perform Membership Inference Attacks (MIAs). Existing defense methods such as confidence masking and generative adversarial networks face practical challenges including decreased model accuracy and high computational cost. This thesis proposes a new defense method based on Moving Target Defense (MTD) to defend against both binary classifier based MIA and metric based MIA. The proposed method uses techniques including perceptual hash, data augmentation, and image fusion to transform training samples, thereby misleading attackers. Additionally, a python script is developed to automatically select appropriate defense intensities for image datasets. Experimental results demonstrate that the proposed defense method is effective and applicable across CIFAR-10, CIFAR-100, Fashion MNIST, Tiny ImageNet and ImageNet-10 dataset under both machine learning and decentralized federated learning pipelines, because the F1-scores of MIAs are reduced to near 0.5 after applying the defense method. Besides, the classification accuracy on test samples is barely changed after applying defense method, indicating that the model performance is not affected for normal users. This makes the method a practical and efficient solution for defending against MIAs.