A Loss-Difference-Based Defense against Membership Inference Attacks

Model overfitting can lead to the success of membership inference attacks, regularization can reduce model overfitting to achieve defense, but this type of defense method tends to affect the model performance. Therefore, this paper proposes a loss-difference-based defense against membership inference attacks. Firstly, the losses are computed for the training and validation sets. Then, the difference between the training loss and validation loss is computed, and compare it with a predefined threshold. If the difference exceeds the threshold, gradient ascent is performed to update the model parameters with the validation set loss. Otherwise, gradient descent is performed to update the model parameters with the training set loss. Experimental results show that the method can significantly reduce the success rate of membership inference attacks while maintaining high model prediction accuracy.