- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper introduces PrivAct, an approach to prevent privacy leakage in LLMs when processing several historical contexts of a user to make inferences of sensitive attributes. The proposed framework is compared to several existing baselines using multiple models.
Critical strengths
The topic that the paper is addressing is quite interesting and relevant; often frontier LLMs start recalling personal information about us when we are chatting with them over multiple sessions, and this could indeed lead them to infer new sensitive attributes by analyzing all these individual disclosures across the context history. I do find this direction valuable; however the paper suffers from clarity and execution issues which I describe in the weaknesses section below.
Critical weaknesses
-
Some parts of writing lack clarity and need further revisions. For example, on line 040 the paper states “The familiar Target pregnancy-prediction case illustrates the underlying aggregation risk outside LLMs Duhigg (2012).”. It seems to be referring to some story as a motivating example but without mentioning more details which for readers that are not familiar, including myself, make it more difficult to understand the point the paper is trying to make. As another example, line 80 of the introduction begins a paragraph that summarizes a lot of results suddenly without prior context, making for a bad reading experience. This also applies to the description of the proposed method in Section 3 which is difficult to fully understand.
-
It is unclear what the takeaways from the results are. Most of the paragraphs describing the results just report the individual numbers of different models with great specificity but do not describe overall trends. There are also too many baselines with no explanations of what they are (e.g. see Table 7), it is difficult to make sense of these results and the paper does not do a good job at it.
Questions for authors
I don't have particular questions, please see the main issues in the weaknesses section.
2 comments
Supplementary Clarification to Reviewer z7dn: Research Motivation and Baseline Comparisons
We sincerely thank the reviewer once again for the valuable suggestions regarding the paper's research motivation, methodological presentation, and interpretation of experimental results. Following our previous responses, we would like to provide two further clarifications.
(1) Regarding the Target example and PRIVACT's research motivation. The New York Times Magazine reported that the retailer Target analyzed patterns in customers' purchases of different products to infer whether they might be pregnant. We cite this example to illustrate that ordinary information that does not directly reveal a sensitive attribute when considered individually may nevertheless disclose a new sensitive conclusion when correlated and combined. This phenomenon is an important motivation for PRIVACT.
Existing representative LLM privacy protection methods primarily focus on identifying known sensitive information, preventing privacy leakage from model memorization, or controlling access to existing sensitive content. However, in multi-turn interactions, LLMs can integrate scattered clues from different historical turns to infer sensitive attributes that were not explicitly disclosed by any individual historical source. This means that privacy risks arise not only from the repeated disclosure of existing sensitive facts, but also from Novel Sensitive Claims formed during the model's reasoning process. PRIVACT is designed to address this problem by combining sensitive claim identification, historical source analysis, and selective protection before the first external release, thereby addressing both the disclosure of existing sensitive information and the formation of new sensitive conclusions.
(2) Regarding the selection of baselines and experimental results. The methods compared in Table 7 represent different technical approaches to LLM privacy protection. For example, DEPN and PATCH primarily focus on sensitive information memorized by models and the mechanisms responsible for its leakage, reducing privacy leakage through model-internal editing. AirGapAgent mainly employs access-control mechanisms, while SumLeaks-CoDef focuses on privacy leakage prevention in multi-agent interactions. Although these methods have different original research objectives and threat models, they represent important technical directions in existing privacy protection research and therefore help evaluate PRIVACT's protection mechanism from different perspectives.
To improve the interpretability of the comparisons, we adapted these methods to a unified evaluation interface and compared them using the same samples and evaluation conditions, while explicitly distinguishing the results of the adapted experiments from the full performance of each method in its original deployment setting. Our central conclusion is not that PRIVACT outperforms existing methods on every metric, but rather that different defense mechanisms exhibit different trade-offs among privacy prevention coverage (Prevention), detection precision (Precision), unnecessary interventions on safe outputs (Safe FI), and task information preservation (Task Preservation). PRIVACT further incorporates historical source novelty into privacy assessment and implements selective protection before the first external release, thereby providing an approach to newly formed sensitive claims in multi-turn interactions that differs from traditional protection against known sensitive information.
We thank the reviewer again for the constructive feedback. We hope that further clarifying the connection between the Target example and our research problem, explaining the technical roles of different baselines, and highlighting the overall trends reflected in the experimental results will make PRIVACT's research motivation, methodological contributions, and experimental findings clearer and easier to understand.
Dear Reviewer z7dn,
Thank you very much for your valuable time and effort in reviewing our paper. Your comments are greatly appreciated and have been helpful in improving the quality of our work.
As the author-reviewer discussion period will end in approximately one day, we would kindly ask you to review our rebuttal and reconsider whether we have adequately addressed your concerns and questions.
Thank you again for your time, consideration, and constructive feedback.
Best regards, The Authors