Source · Tier A · Peer-reviewedSource · Hollow-LLM Attack: Computationally Trivial Weights in Zero-Knowledge Verification of LLM Inference
Hollow-LLM Attack: Computationally Trivial Weights in Zero-Knowledge Verification of LLM Inference
C. Gong, B. Liu, M. Li. 2026. 2026 IEEE Symposium on Security and Privacy (SP).
| Original | https://arxiv.org/abs/2607.28884 |
|---|---|
| DOI | 10.1109/SP63933.2026.00258 |
| arXiv | 2607.28884 |
| Version | Published in the 2026 IEEE Symposium on Security and Privacy (SP), pp. 4339–4355. Crossref, checked on 2026-10-17, gives an issue date of 18 May 2026 and the same author order as the paper's title page. The text read is arXiv v1 (30 July 2026), via its HTML and PDF renderings. Its comments field says it was accepted to IEEE S&P 2026. |
| Accessed | 2026-09-25 |
| Note | Independent analysis (University of Southern California) of zero-knowledge proofs of LLM inference. Shows that valid proofs do not bind the computation spent, using "ghost weights"; experiments use the proof procedure of zkGPT on a 6-layer, 512-dimensional transformer. |
Cited by
- Zero-knowledge proofs of inferenceMechanism
- zkLLMImplementation