Reminiscence Attack on Residuals: Exploiting Approximate Machine Unlearning for Privacy
Yaxin Xiao [email protected], Qingqing Ye [email protected], Huadi Zheng [email protected], Haibo Hu [email protected], Zi Liang [email protected], Haoyang Li [email protected], Yijie Jiao [email protected], Department of Electronical and E · 2025
Read Reminiscence Attack on Residuals: Exploiting Approximate Machine Unlearning for Privacy by Yaxin Xiao [email protected], Qingqing Ye [email protected], Huadi Zheng [email protected], Haibo Hu [email protected], Zi Liang [email protected], Haoyang Li [email protected], Yijie Jiao [email protected], Department of Electronical and E (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-20573.txt.
Description
Yaxin Xiao and co-authors show that teaching a model to forget data can leave traces an attacker can read, in a 2025 paper. Machine unlearning removes specific data so people can exercise a right to be forgotten, and fast approximate methods avoid full retraining. The team found these methods leave implicit residuals, faint imprints of the removed data, whatever the architecture or algorithm used.
Their Reminiscence Attack amplifies those residuals with targeted fine-tuning, then infers whether data was in the training set, beating earlier attacks by wide margins. As a defense, they built a two-phase unlearning method that scrubs deep-layer traces and stabilizes training, cutting attack accuracy to near guessing while keeping the forgetting effective. The work reframes unlearning as a privacy problem of its own.
Catalog record
| CREATOR | Yaxin Xiao [email protected], Qingqing Ye [email protected], Huadi Zheng [email protected], Haibo Hu [email protected], Zi Liang [email protected], Haoyang Li [email protected], Yijie Jiao [email protected], Department of Electronical and E |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 73 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-20573