Robustifying ML-powered Network Classifiers with PANTS#
USENIX Security '25 (34th USENIX Security Symposium) digest Program Core pick 13–15 August 2025
Minhao Jin and Maria Apostolaki, Princeton University — Core
- Why: ML-based traffic classification is increasingly load-bearing for resource allocation and intrusion detection in cloud-native networking, and this is a practical, white-box framework for finding and fixing adversarial blind spots in exactly that class of system.
- What: PANTS, a white-box framework combining adversarial ML with SMT solvers to generate realizable adversarial inputs against ML-based network traffic classifiers (MNCs), plus an iterative adversarial-training process that hardens MNCs against them. Outperforms prior baselines (Amoeba, BAP) by 70%–2x in finding adversarial inputs and improves target-model robustness by 52.7% without sacrificing accuracy. (from abstract only)
- How: Integrates SMT solving to handle MNC-specific challenges — non-differentiable components like traffic engineering, and semantic/reliability constraints on realizable inputs — that block direct use of standard gradient-based adversarial ML methods. (from abstract only)
- Where applicable: General for network operators deploying ML-based traffic classification (intrusion detection, resource allocation); a practical robustness-testing and hardening tool, not just an attack demonstration.