Skip to main content
Back to timeline
arXivSource publication:

A legitimate deletion request can erase an unnamed entity: proxy unlearning turns shared representations into a targeted threat

Synopsis

Large language models encode distinct entities in shared parameter subspaces, and this knowledge superposition lets a legitimate unlearning request degrade sensitive entities that were never requested while removing the entities that were. The researchers name this exploitation the Proxy Unlearning Attack (PUA) and propose Superposition-Aware Unlearning Protection (SAUP), which detects at-risk entities at the representation level before unlearning and, afterward, decouples shared representations and restores them while keeping the requested entities suppressed. Across three knowledge domains and two model families, PUA produces selective target degradation and SAUP recovers most of the damaged knowledge.

Source-provided article image: Protecting Sensitive Knowledge Under LLM Unlearning: The Proxy Threat and a Superposition-Aware Defense
arXiv

Interpretation

A fully legitimate unlearning request can visibly degrade a model's knowledge of a sensitive entity without ever naming that entity. Earlier unlearning work mainly asked whether requested entities were removed and general utility preserved, treating collateral damage as a utility or evaluation concern; here collateral damage is treated as an exploitable targeted security risk. On LLaMA-3.1-8B and the Qwen-2.5 family across CounterFact, UMLS and SBIC, the target group was never submitted for removal yet its Judge score fell alongside the proxy group; on LLaMA SBIC under GA it dropped from 0.850 to 0.281.

An attacker only needs to submit, through the ordinary unlearning interface, proxy entities whose representations overlap the target, concentrating the harm without making the request look suspicious. Unlike attacks that require prior poisoning, backdoor planting, or comparing original and unlearned models, this attack needs no prior access to training data or model internals; the harm comes from correct execution of unlearning itself. Against a request-size-matched random-proxy request, KSD-selected proxies lowered the target by 0.331 on average versus 0.052 for random requests; on Qwen, black-box transfer from a same-family 3B surrogate still lowered the target by 0.090 to 0.168.

A provider can add one step before and one after unlearning, without modifying the unlearning algorithm, to honor the deletion and restore the protected entities that were collaterally weakened. Naive fine-tuning recovery restores the requested entities too, because they share representation directions, while output-level post-hoc detection only sees already visibly degraded samples. On LLaMA, SAUP raised the target and kept a positive target–proxy gap in all six cells, reaching 0.460 versus 0.781 on SBIC under GA+SAUP; against an output-level detection-and-recovery baseline, SAUP kept a wider target–proxy gap in all three domains.

Perspective

Evaluation is scoped to entity-level factual knowledge, scored on definition, category and association probes, and does not cover general task ability; experiments run locally on public open-weight models, LLaMA-3.1-8B-Instruct and the Qwen-2.5 family, over the CounterFact, UMLS and SBIC knowledge domains.

Recovering protected entities can also raise the requested proxies: on LLaMA CounterFact the proxy score rises by 0.213 under GA and 0.291 under NPO after recovery, a measurable trade-off between protecting targets and strictly completing the deletion request.