Quantifying the Privacy Impact of High-Fidelity Synthetic Network Traffic

The researchers propose a privacy measurement framework for synthetic network traffic that combines membership inference attacks (MIA), data extraction attacks, and network-specific identifiers to systematically evaluate several representative generative models. Risk varies significantly across models and datasets: MIA success rates range from 0% to 88%, and up to 100% of network identifiers can be recovered from generated traffic. Training data diversity and the degree to which models fit the training data significantly affect attack outcomes, providing guidance for designing generative models with low privacy leakage.