CertTA: Certified Robustness Made Practical for Learning-Based Traffic Analysis
No ratings
Presented at USENIX Security 2025 by
Learning-based traffic analysis models exhibit significant vulnerabilities to adversarial attacks. Attackers can compromise these models by generating adversarial network flows with precisely optimized perturbations. These perturbations typically take two forms: additive modifications, which include packet length padding and timing delays, and discrete alterations, such as dummy packet insertion. In response to these threats, certified robustness has emerged as a promising methodology for ensuring reliable model performance in the presence of adversarially manipulated network traffic.