CoReg: Membership Privacy Protection via Collaborative Regularization

Machine learning is widely used in many areas, however, it has been proved that machine learning models are prone to leak sensitive information of the training data. For instance, the adversary could precisely infer whether a sample belongs to the target model's training set with membership inference attacks (MIAs). To mitigate the membership inference risks, we propose a novel defense framework, named CoReg. It trains multiple sub-models and they regularize each other with a novel collaborative regularization. With this method, we could reduce the membership leakage risks by enforcing the model to have similar behavior on members and non-members, while maintaining high classification performances. Furthermore, we propose an adversarial output control module to select the output of the least risky sub-model as the final output. We carry out extensive experiments on three datasets and verify that CoReg could achieve better protection effects against MIAs than baselines while maintaining high classification accuracy.