FINDING · DETECTION
Joint multi-task training with a combined loss L_joint = L_site + λ·L_pers shows that increasing λ from 0 to 2 raises mixed-site persona accuracy from approximately 45% to approximately 80% while website accuracy declines only from approximately 90% to approximately 75%, demonstrating a wide regime where an attacker can gain strong persona inference at modest cost to existing WFP capability.
From 2026-song-personafingerprint-measuring-persona — PersonaFingerprint: Measuring Persona Inference on Modern Websites with LLM-Driven Browsing · §4.4, §5.7.2, Figure 5 · 2026 · arXiv preprint
Implications
- An attacker with access to persona-labeled traffic (e.g., from LLM-driven browsing agents) can upgrade a deployed WFP pipeline to dual-purpose persona inference at low marginal cost — circumvention deployments must not assume that site-level defenses also bound persona-inference risk.
- Circumvention protocol evaluations should include dual-head adversary models that jointly optimize site and behavioral classification, since these represent realistic near-future attack capabilities.
Tags
Extracted by claude-sonnet-4-6 — review before relying.