Toward Privacy-Preserving Training of Generative AI Models for Network Traffic Classification

Synthetic traffic traces are useful for training traffic classifiers in privacy-constrained environments. Generative Artificial Intelligence (GAI) models are blossoming as a solution to avoid the sharing of real data and the lack of datasets. Never the less ,privacy concerns about GAI are often under-estimated. Therefore, an approach to mitigate the data leakage of a GAI is presented in this paper, with a minimum impact on the utility of synthetic traffic traces for downstream applications. For example, training a Machine Learning (ML) traffic classifier on synthetic traffic traces results in an average accuracy loss of no more than 13% concerning training on a real dataset.