Redefining Website Fingerprinting Attacks with Multi-Agent LLMs#

Authors: Chuxu Song (Rutgers University), Dheekshith Dev Manohar Mekala (Rutgers University), Hao Wang (Rutgers University), Richard P. Martin (Rutgers University)

• Why: Two fields you track — network traffic analysis and LLM agents — collide here. If multi-agent LLM pipelines can replace hand-crafted website-fingerprinting feature engineering, the barrier to sophisticated traffic analysis drops dramatically, and the same shift threatens every hand-tuned classifier in your security stack.
• What: Apparently a re-formulation of website fingerprinting (inferring visited sites from encrypted traffic traces) where multi-agent LLM systems perform the analysis instead of fixed ML models (title/author line only — no abstract in program).
• How: Inference: traffic traces are serialized into representations that LLM agents reason over, possibly with role-specialized agents (feature extraction, pattern matching, decision) orchestrated together.
• Where applicable: Traffic-analysis research and, adversarially, any encrypted-traffic classification problem — VPN/anonymity users today, but the methodology generalizes to flow-based monitoring generally. Compute cost likely limits real-time use for now.