ReplayAudit audits outsourced LLM fine-tuning via sealed records and selective replay
Original title: ReplayAudit: Auditing Outsourced LLM Fine-Tuning via Sealed Records and Selective Replay
The system uses telemetry-guided sampling to audit training with 90% probability of detecting injected steps while reducing audit time by up to 67%.
ReplayAudit is a telemetry-guided auditing system for outsourced LLM fine-tuning. It seals training records before challenges and replays sampled intervals from sealed checkpoints. Loss changes and parameter-update norms guide sampling. Evaluation on real LoRA training shows it rejects false reports and reduces wall time compared to uniform sampling.
- 7B BF16 prototype rejects false update-norm reports
- calibrates sampling budgets offline to at least 90% probability of directly selecting an injected step
- reduces median complete-audit wall time by 49% and 67% for two learning-rate deviation settings
- when skipped updates evade, audit costs 38% more
- AI models
- LLM
From the original article
1 min read at arxiv.orgAuditing outsourced LLM fine-tuning requires checking an untrusted provider's training while controlling verification cost. Uniform sampling ignores training-time signals, while checking a suspicious record can require replaying several preceding steps.
By Yihan Xia · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.