GANDaLF: GAN for Data-Limited Fingerprinting
Published in Proceedings on Privacy Enhancing Technologies (PoPETS), 2021
In this paper, we introduce Generative Adversarial Networks for Data-Limited Fingerprinting (GANDaLF), a deep learning-based approach designed to perform Website Fingerprinting (WF) on Tor traffic. Unlike previous WF attacks that require large training datasets, GANDaLF uses a GAN to generate fake training samples to improve classifier performance with limited real data. Our experiments show that GANDaLF achieves 87% accuracy with only 20 training samples per site in a closed-world scenario. Furthermore, it significantly outperforms prior state-of-the-art methods, particularly in subpage fingerprinting, where GANDaLF achieves a 29% improvement over Triplet Fingerprinting (TF) and a 38% improvement over Var-CNN.
Recommended citation: Se Eun Oh, Nate Mathews, Mohammad Saidur Rahman, Matthew Wright, Nicholas Hopper. (2021). "GANDaLF: GAN for Data-Limited Fingerprinting." Proceedings on Privacy Enhancing Technologies (PoPETS), 2021. DOI: 10.2478/popets-2021-0029.
Download Paper | Download Slides