VAE-MIP: A Mutual Information Constrained Variational Autoencoder for Privacy Protection Data Generation

Existing privacy protection methods primarily reduce sensitivity to training data by adding noise during the training process or constraining model attributes. These methods struggle to strike a balance between the strength of privacy protection and the utility of data. Although common privacy protection methods based on differential privacy or adversarial generative models are widely used to enhance privacy protection strength, it's struggle to maintain high-quality data generation while improving privacy protection quality. In this paper, we propose a variational autoencoder based on minimizing mutual information constraints, namely VAE-MIP. VAE-MIP enhances privacy protection by introducing a mutual information penalty term into the encoder's loss function Additionally, the CBAM attention module is integrated into the model to improve the quality of generated data. Experiments on the CIFAR-10 and MNIST datasets demonstrate that the proposed method outperforms baseline models in both privacy protection strength and the quality of generated data.