- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The authors argue that text anonymizers are mostly evaluated in terms of adversarial accuracy, the share of attributes an LLM attacker can still infer, with many of the approaches using the same attacker within their rewriting looping. A low accuracy may therefore mean that the attacker was misled and not that the attribute is removed. To test this, a linear probe on a small open model's hidden states is used to score author-level AUC. Empirical results show that the two measures diverge, i.e., the anonymizers cut attacker accuracy significantly, yet frozen-probe AUC stays close the raw text. PGRS, the proposed defense, generates 8 register-directed rewrites and publishes the one with the highest frozen-probe entropy. With this defense, AUC drops significantly.
Critical strengths
- Proposes a new well-motivated evaluation measure with empirical demonstrations. The authors provide convincing evidence for the gap in evaluating text anonymizers only in terms of adversarial accuracy. For example, IncogniText illustrates how adversarial accuracy can mislead: roughly half of the attacker's first guesses hit the planted wrong value, driving accuracy below the majority prior while probe AUC remains near its raw-text level. RLAA shows the same gap with minimal edits. Extending Elazar & Goldberg (2018) to text anonymization, the paper offers an actionable recommendation to report both frozen and retrained probe AUC.
Critical weaknesses
- PGRS's practical advantage is unclear. The authors acknowledge that frozen-probe AUC overstates protection and that retrained-probe AUC is the meaningful test. However, the ablations favor a simpler alternative: applying one fixed register to every comment achieves retrained AUC 0.630 at utility 0.868, versus PGRS's 0.582 at utility 0.758. Probe-guided selection thus gains only 0.05 AUC at a cost of 0.11 utility. The claim that PGRS improves protection "without a loss on adversarial accuracy or utility" holds only against FgAA (e.g., its utility is substantially below RLAA). Moreover, the retrained result may be optimistic: both probes are trained on the same development authors, allowing the adversary to learn from outputs selected in-sample to confuse a probe trained on those authors.
Questions for authors
Q. Given Table 3, why should practitioners prefer full PGRS over fixed-register rewriting, which is cheaper and has higher utility at a similar retrained AUC? Can you show the privacy-utility frontier under the retrained probe?