- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes three practical mechanisms for post-hoc MIA-based privacy auditing of LLMs and VLMs. It reduces distributional shift between member and non-member examples by rewriting non-member texts and filtering both groups for greater empirical overlap, and improves MIA performance by adding paraphrase-related features from auxiliary LLMs. The experimental results across multiple datasets and models show effective privacy leakage estimates with reduced distributional shift and a retained MIA signal with these techniques.
Critical strengths
- Practical improvements to post-hoc privacy auditing: The paper proposes non-member rewriting, overlap-based filtering, and local-paraphrase-based features for MIA to reduce data-only distinguishability between member and non-member data while retaining MIA performance, providing a practical privacy auditing approach for LLMs and VLMs.
- Extensive experiments and ablation studies: The experiments cover a wide range of target models and datasets, providing evidence for the effectiveness and generalizability of the proposed techniques. The ablation studies also demonstrate the contribution of each component to the claimed improvements.
Critical weaknesses
- Residual distributional shift may be mistaken for privacy leakage: In this work, BERT models are used to encode text inputs for the discriminator, while the target models are LLMs/VLMs. The discriminator may miss distributional shift that the LLM-/VLM-based model can exploit. It remains unclear whether the performance gap reflects actual privacy leakage or the residual distributional shift that the discriminator fails to capture.
- Inconsistent result reporting: Some reported results seem to be inconsistent across tables and sections, leading some quantative comparisons and conclusions unclear. For example:
- Table 3 reports values ranging from 0.42 to 0.60 for OpenClip models on LAION and cc12m, which are lower than the reported PANORAMIA estimates, while Section 3.1 states them as 1.74, 2.09, and 2.06, and are larger instead.
- Table 3 reports as 1.72 for Gutenberg/OpenLLaMA-3B, while Section 3.1 states it as 1.71.
- Section 3.2 states that is the largest for LAION benchmark when filtering at , but Table 4 shows the largest value at (1.7040 at vs. 1.4776 at ).
- Table 3 reports TPR@10%FPR as 0.5994 for LAION/Vit-B-32, while the value is 0.5594 in Table 4 for the same setup.
Questions for authors
- Conflicting reported results: Which reported are correct for the OpenClip models on LAION and cc12m? Could the authors clarify the discrepancy between Table 3 and Section 3.1? Would the correction change the interpretation of privacy leakage estimates in Section 3.1?
- WikiMIA auditing results: In Table 3, the MIA yields a lower TPR@k%FPR than the discriminator across all target models while providing positive privacy estimates. Section 3.1 attributes this to the better performance at very low FPR, which is absent in the table. Could the authors report these results to support the claim? Is there any explanation or intuition behind this performance trend?
- Effect of rewriting on MIA performance: Why does rewriting increase MIA's TPR@k%FPR for Temporal-Wiki/Pythia-2.8B in Table 4? Does this reflect better separation of member and non-member data, or artifacts introduced by rewriting?
3 comments
Reviewer interaction needed
Dear reviewer,
The authors have replied to your comments and the discussion is close to the end. We encourage you to participate in the discussion to see whether your questions and concerns have been answered by the authors. Thank you!
Best, Your AC
Thank you for the clarifications and corrections to the experimental results. My concerns have been addressed: the additional GPT-2 experiment helps address my concern about the discriminator, the reporting inconsistencies have been corrected, and the explanations of the WikiMIA and rewriting results are helpful.
I will raise my score from 2 to 3. Thanks!
W1: “..The discriminator may miss distributional shift that the LLM-/VLM-based model can exploit..”
To directly investigate whether this difference in feature information explains the observed membership advantage, we conduct an additional experiment on Temporal-Wiki using GPT-2 as an independent reference model. GPT-2 was pretrained on WebText, which explicitly excluded Wikipedia pages, making it a suitable reference model with different training provenance from the audited Pythia models. We extract the same paraphrase-based z-score features from GPT-2 as those used in our target-model MIA and incorporate them into the data-only discriminator. This allows the discriminator to capture additional distributional differences detectable through autoregressive language-model features, without accessing the audited target model. As shown in Section 3.2 Table 7, the adjusted leakage estimate remains positive after adding these reference-model features. Moreover, the observed lower bound on the c_closeness remain almost identical for two of the reference model and comparable (well within one std) for the largest reference model. This provides evidence that the MIA's advantage is not simply due to its access to paraphrase-based likelihood features, but reflects additional membership information obtained specifically from the audited target model. For other benchmarks, identifying a suitable autoregressive reference model with documented non-overlap with the audited member records is considerably more difficult. A reference model that has itself trained on the audited members could introduce membership information into the discriminator, complicating its interpretation as a data-only baseline. We therefore retain the BERT-based discriminator for these settings and acknowledge this limitation. For the VLM experiments, our existing discriminator already incorporates features from a pretrained reference CLIP model, including cross-modal representations and similarity scores comparable to those used by the target-model MIA (Figure 3). The resulting leakage estimates remain positive, providing additional evidence that the observed membership advantage is not simply attributable to the target-model MIA having access to richer feature representations.
W2&Q1: “Inconsistent result reporting..”
We apologize for these errors in writing, which we have corrected in the paper. Table 3 contains the correct VLM results. We have updated Table 3 with the addition of standard deviations. We have corrected the numerical values and corresponding discussion in Section 3.1 to match Table 3. We note that it does not change our main conclusion: we do detect positive privacy leakage in these settings.
Q2: “WikiMIA auditing results..”
For that particular run of WikiMIA we report [email protected]%FPR scores for discriminator and MIA as follows: Discriminator: 0.0, Pythia2.8B/12B: 0.0625, Pythia6.9B: 0.0129. This distinction at smaller FPR rate is relevant for WikiMIA because of its small sample size. The benchmark contains only 388 members and non-members each, and is further reduced to 155 per group after filtering. At this size, a single false positive already corresponds to approximately 0.65% FPR, so the reported TPR@1%FPR provides limited resolution of the extreme tail. We have updated Table 7 with mean and standard deviation over 5 runs with random seed and the mean values behave more robustly as expected.
Q3: “Effect of rewriting on MIA performance..”
The increase in TPR after rewriting should be interpreted in the context of the changed non-member distribution, and we do not interpret the raw increase in MIA TPR as evidence of stronger membership leakage. Since TPR is measured at a fixed FPR with respect to the non-member data, rewriting changes the empirical data distribution of non-members and the learned classification boundary and affects TPR as well. In this specific case, we observe that rewriting shifts non-member MIA scores downward, and the decision threshold corresponding to 1% FPR also decreases, allowing more member data to be correctly classified and increasing TPR. Importantly, our discriminator is designed to capture these distributional differences observable from the data, including potential artifacts introduced by rewriting. Such differences are accounted for in our discriminator-adjusted audit.