<p>Membership Inference Attacks (MIAs) can exploit machine learning model prediction interfaces to infer sensitive information about the training data, posing serious risks to the privacy of the training data. A recent study explored knowledge distillation as a potential method to improve model robustness against MIA by transferring knowledge from a teacher model trained on sensitive data to a target model using public data, thereby reducing the target model’s reliance on private training data. However, this approach typically adopts identical architectures for both teacher and target models, which limits the teacher’s ability to provide diverse representations, impairing the performance and privacy resilience of the target model. To this end, we propose a novel defense scheme that uses the CLIP-Conditioned Distillation method for privacy protection, called C2DP. C2DP employs the image encoder of CLIP with an added classification head as a high-capacity teacher model, which enhances knowledge transfer while avoiding privacy risks caused by architectural homogeneity. During distillation, the teacher generates soft and pseudo-hard labels from a public dataset, which provide complementary supervision signals to improve training efficiency and generalization of the target model. Extensive experiments on real-world datasets show that the proposed scheme achieves superior model utility and enhanced protection against Membership Inference Attacks compared to representative approaches.</p>