Evidence · DiggingBeagle record
TRACE: training-data reconstruction from unlearning checkpoint differences
TRACE: training-data reconstruction from unlearning checkpoint differences
- Source role
- primary disclosure
Evidence record
TRACE: training-data reconstruction from unlearning checkpoint differences
Claim-level citations (2)
- supportsBefore-and-after weights can expose data an unlearning procedure was meant to erase: TRACE assumes the attacker can obtain model snapshots from both before and after approximate unlearning, then uses embedding differences and a perplexity-gap search to reconstruct removed samples without possessing the original training examples.
Abstract, before-and-after checkpoint access and reconstruction steps
- supportsBefore-and-after weights can expose data an unlearning procedure was meant to erase: TRACE demonstrates reconstruction against selected approximate-unlearning methods using paired pre-removal and post-removal checkpoints, with benchmark examples rather than a reported victim data leak.
Abstract; two-checkpoint attack assumptions; TOFU, MUSE and WMDP results