Preprint · 2025 · Open access
PrivacyScalpel: Enhancing LLM Privacy via Interpretable Feature Intervention with Sparse Autoencoders
arXiv (Cornell University) · Published 14 March 2025 · DOI 10.48550/arxiv.2503.11232
In brief
“PrivacyScalpel: Enhancing LLM Privacy via Interpretable Feature Intervention with Sparse Autoencoders” is a 2025 preprint by Frikha, Ahmed and 5 co-authors, published in arXiv (Cornell University). OpenAlex counts 8 citations as of 19 November 2026. It is open access: the full text is free to read at arXiv (Cornell University).
- Citations
- 8
- Most cited in
- 2025 (4)
- References
- 27
- Authors
- 6
Abstract
Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing but also pose significant privacy risks by memorizing and leaking Personally Identifiable Information (PII). Existing mitigation strategies, such as differential privacy and neuron-level interventions, often degrade model utility or fail to effectively prevent leakage. To address this challenge, we introduce PrivacyScalpel, a novel privacy-preserving framework that leverages LLM interpretability techniques to identify and mitigate PII leakage while maintaining performance. PrivacyScalpel comprises three key steps: (1) Feature Probing, which identifies layers in the model that encode PII-rich representations, (2) Sparse Autoencoding, where a k-Sparse Autoencoder (k-SAE) disentangles and isolates privacy-sensitive features, and (3) Feature-Level Interventions, which employ targeted ablation and vector steering to suppress PII leakage. Our empirical evaluation on Gemma2-2b and Llama2-7b, fine-tuned on the Enron dataset, shows that PrivacyScalpel significantly reduces email leakage from 5.15\% to as low as 0.0\%, while maintaining over 99.4\% of the original model's utility. Notably, our method outperforms neuron-level interventions in privacy-utility trade-offs, demonstrating that acting on sparse, monosemantic features is more effective than manipulating polysemantic neurons. Beyond improving LLM privacy, our approach offers insights into the mechanisms underlying PII memorization, contributing to the broader field of model interpretability and secure AI deployment.
Cite this paper
Formatted from the catalogue record. Check names, capitals and page numbers against the paper before you submit.
APA 7
Ahmed, F., Ar, R. M. R., Nakka, K. K., Mendes, R., Jiang, X., & Zhou, X. (2025). PrivacyScalpel: Enhancing LLM Privacy via Interpretable Feature Intervention with Sparse Autoencoders. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2503.11232
MLA 9
Ahmed, Frikha,, et al. “PrivacyScalpel: Enhancing LLM Privacy via Interpretable Feature Intervention with Sparse Autoencoders.” arXiv (Cornell University), 2025. https://doi.org/10.48550/arxiv.2503.11232.
Chicago
Ahmed, Frikha,, Razi, Muhammad Reza Ar, Krishna Kanth Nakka, Ricardo Mendes, Xue Jiang, and Xuebing Zhou. 2025. “PrivacyScalpel: Enhancing LLM Privacy via Interpretable Feature Intervention with Sparse Autoencoders.” arXiv (Cornell University). https://doi.org/10.48550/arxiv.2503.11232.
Harvard
Ahmed, F. et al. (2025) ‘PrivacyScalpel: Enhancing LLM Privacy via Interpretable Feature Intervention with Sparse Autoencoders’, arXiv (Cornell University). doi:10.48550/arxiv.2503.11232.
IEEE
F. Ahmed, R. M. R. Ar, K. K. Nakka, R. Mendes, X. Jiang, and X. Zhou, “PrivacyScalpel: Enhancing LLM Privacy via Interpretable Feature Intervention with Sparse Autoencoders,” arXiv (Cornell University), 2025, doi: 10.48550/arxiv.2503.11232.
BibTeX
@misc{ahmed2025privacyscalpel,
title = {{PrivacyScalpel: Enhancing LLM Privacy via Interpretable Feature Intervention with Sparse Autoencoders}},
author = {Ahmed, Frikha, and Ar, Razi, Muhammad Reza and Nakka, Krishna Kanth and Mendes, Ricardo and Jiang, Xue and Zhou, Xuebing},
howpublished = {arXiv (Cornell University)},
year = {2025},
doi = {10.48550/arxiv.2503.11232},
url = {https://doi.org/10.48550/arxiv.2503.11232}
}
Download for a reference manager: BibTeX (.bib) · RIS for Zotero, Mendeley and EndNote (.ris)
Topic and keywords
Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
- sparse autoencoders
- ablation
- interpretability
- privacy-utility trade-off
- privacy-preserving AI
Work with this paper
Most-cited papers in Computer Science
- ImageNet classification with deep convolutional neural networksAlex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton · Communications of the ACM · 2017 · 114,726 citations
- Long Short-Term MemorySepp Hochreiter, Jürgen Schmidhuber · Neural Computation · 1997 · 109,097 citations
- A Mathematical Theory of CommunicationClaude E. Shannon · Bell System Technical Journal · 1948 · 91,736 citations
- Deep learningYann LeCun, Yoshua Bengio, Geoffrey E. Hinton · Nature · 2015 · 90,226 citations
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal NetworksShaoqing Ren, Kaiming He, Ross Girshick and 1 more · IEEE Transactions on Pattern Analysis and Machine Intelligence · 2016 · 74,143 citations
Questions about this paper
How do I cite this paper in APA style?
Ahmed, F., Ar, R. M. R., Nakka, K. K., Mendes, R., Jiang, X., & Zhou, X. (2025). PrivacyScalpel: Enhancing LLM Privacy via Interpretable Feature Intervention with Sparse Autoencoders. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2503.11232
How many times has this paper been cited?
8 times, according to OpenAlex on 19 November 2026. Google Scholar and Scopus count different sets of sources, so their numbers differ.
Is this paper free to read?
Yes. It is open access at arXiv (Cornell University): sign in and open the PDF from this page.
Who wrote this paper?
Frikha, Ahmed, Razi, Muhammad Reza Ar, Krishna Kanth Nakka, Ricardo Mendes, Xue Jiang, Xuebing Zhou.
What is the DOI of this paper?
10.48550/arxiv.2503.11232, permanently at https://doi.org/10.48550/arxiv.2503.11232. Sign in and open it from this page.
Bibliographic data from OpenAlex (CC0), last checked 19 November 2026. ResearchGuru does not host this paper; once you are signed in, the links open it at the publisher or the repository that does.