Privacy-Preserving Wi-Fi Data Generation via Differential Privacy in Diffusion Models

Due to the considerable effort and time needed to collect and label wireless data, there is a compelling need for data generation to facilitate data augmentation. To ensure the reliability of the data, the generated data needs to perform well in common evaluation metrics. However, this process can lead to the model memorizing some training data, resulting in potential privacy leaks. One major threat is the membership inference attack (MIA), which determines whether a specific sample was used in training the target model. While MIA has been extensively studied for discriminative models, its impact and defenses for generative models remain less explored. In this paper, we propose a hybrid training method for the diffusion model applied to Wi-Fi data as a defense against MIAs. The approach involves initially training the model without privacy constraints. After a specified number of training rounds, differential privacy (DP) is incorporated for fine-tuning. During this second phase, a co-optimization process is conducted in parallel to counteract the effects of the added noise. Experimental results demonstrate that the hybrid training method effectively defends against state-of-the-art MIAs for generative models without compromising model performance or requiring additional training efforts, showing significant promise for practical applications.