Efficient Techniques for Data Reconstruction, with Finite-Width Recovery Guarantees
Edward Tansley, Roy Makhlouf, Estelle Massart, and, Coralia Cartis · 2026
Read Efficient Techniques for Data Reconstruction, with Finite-Width Recovery Guarantees by Edward Tansley, Roy Makhlouf, Estelle Massart, and, Coralia Cartis (2026) 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-2605-06519.txt.
Description
Edward Tansley, Roy Makhlouf, Estelle Massart, and Coralia Cartis examine attacks that reconstruct a neural network's training data from its weights, a serious privacy threat when data are sensitive. They propose one unified optimization formulation based on initial and trained parameter values, gathering earlier proposals into a single frame.
For random feature models they prove reconstruction succeeds with high probability when the network is wide enough, a first finite-width guarantee of this kind. If data lie in a low-dimensional subspace, the needed width drops. Their practical algorithm estimates that subspace from first-layer weight changes and reconstructs using only last-layer weights, beating full-space methods in tests.
Catalog record
| CREATOR | Edward Tansley, Roy Makhlouf, Estelle Massart, and, Coralia Cartis |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 54 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.06519