TRACE: A Data-Free Text Reconstruction Attack Against Approximate Unlearning in Large Language Models
TRACE: Data-Free Text Reconstruction Attacks against Approximate Unlearning in LLMs
Translation in progress — untranslated text is omitted.
Researchers propose the TRACE attack framework, which uses only model snapshots before and after unlearning to reconstruct approximately unlearned training data in large language models, without any data-specific prior. The method filters candidate tokens from parameter differences at the embedding vector level to narrow the search space, then assembles sequences using a contrastive scoring mechanism based on the perplexity gap between the two models, and recovers multiple distinct samples through a two-stage explore-then-complete strategy. Experiments on the TOFU, MUSE, and WMDP datasets, multiple large language models, and six representative unlearning methods show that TRACE achieves high-fidelity reconstruction. The paper points out that current LLM approximate unlearning deployments pose significant privacy risks, and defenses against parameter-difference leakage are needed along with proper management of model snapshot access.
- LLM deployments often retain model snapshots from before and after unlearning for auditing and rollback; the authors ask whether approximate unlearning leaves recoverable traces in the snapshot parameter differences, allowing deleted data to be reconstructed without any data priors.
- Proposes TRACE: use embedding-layer parameter differences to filter a candidate token set and compress the search space, use a contrastive fluency score (CFS) based on the perplexity difference between the two models to guide the search, then use a two-stage "broad prefix exploration—prefix completion" process to recover multiple different samples.
- Evaluated on TOFU, MUSE, and WMDP across 6 models and 6 unlearning methods: on TOFU, LLaMA2-7B after GA unlearning reaches A-ESR1.0 of 95.00%; on WMDP long text, the authors state that ROUGE-L(R) is more than double that of UBNF with a 10-word prefix; it remains partially effective under black-box, quantization, and pruning.
- Limitations listed by the authors: white-box parameter access to both snapshots is required; the method relies on embedding-layer parameter differences, and future methods that specifically hide this signal may affect effectiveness; reconstruction quality is affected by model memorization strength and unlearning batch size. Experiments use only open-source models, and attack runtime and query counts are not reported.
Related papers
- Test-Time Unlearning via Sparse AutoencoderDefensesarXiv