FINDING · DETECTION
State-of-the-art ML-based obfs4 detection (Wang et al. decision tree) achieves 97% precision at equal base rates (λ=1) but precision collapses to 3% at a still-conservative λ=1,000; at λ=10⁶ precision approaches zero for all classifiers tested. This base-rate failure was previously uncharacterized because prior evaluations only considered balanced or near-balanced datasets.
From 2024-wails-precisely — On Precisely Detecting Censorship Circumvention in Real-World Networks · §IV-D (Scalability), Figure 3, Table I · 2024 · Network and Distributed System Security
Implications
- Circumvention protocol evaluations must report Precλ across at least six orders of magnitude of base rate, not just λ=1, to expose false-positive collapse that makes flow-level blocking infeasible at scale.
- Hiding in the long tail of rarely-seen destination ports raises the censor's collateral-damage cost per true positive, but does not prevent a well-calibrated host-level classifier from eventually identifying proxy servers.
Tags
Extracted by claude-sonnet-4-6 — review before relying.