Measuring re-execution checks and attack detection for LoRA training on Apple silicon
Training verification work funded here this year targets data centres: zero-knowledge proofs of training properties (arXiv 2606.05433) and workload classification from hardware sensors (arXiv 2606.19262). This project measures the cheap route on the other kind of hardware: can a fine-tuning run on an untrusted consumer Mac be checked by re-running a sample of it, and what does a cheating machine owner get past that check? The trainer already exists and is exactly repeatable on one machine. The funding pays for the verification experiments, not for building the fine-tuning system.
Three experiments over twelve weeks. First, run one training job on several Apple chip generations and record whether the adapters match bit for bit, and if not, at which step they part and by how much. Second, build a verifier that re-runs a random sample of training segments between checkpoints and compares hashes, and measure its cost as a share of the original run. Third, attack it: swap training rows to plant a backdoor, skip steps, substitute the base model, and measure the detection rate for each attack at each sampling rate. Re-execution across hardware has been shown for three NVIDIA GPUs (Srivastava, Arora and Boneh, 2024). Zero-knowledge proofs for LoRA are reported at 17 to 31 seconds per training step. I found nothing published for Apple silicon, which is what consumer inference networks run on.
A public report with three numbers: the step at which cross-chip runs diverge (or confirmation that they don't), the verifier's cost as a percentage of training, and a detection rate per attack. Code and raw logs are released with it.
At the $3,000 minimum: a rented Mac mini M6 for three months ($650), a second rented Apple chip of another generation (about $650), a Claude subscription for three months ($600), and rented NVIDIA time to reproduce the published baseline (about $950). With my own M1 Pro that gives three chips. At $10,000: adds a 24 GB MacBook Pro M5 as a fourth chip and the only machine I can run long attack jobs on without renting (about $1,800), and three months of my time ($5,200). I have a pending BlueDot Rapid Grant application for $5,050 that overlaps the M6 rental, Claude and laptop. If it is approved I will lower this goal by the overlap and note it here.
I work alone. I am a fellow at Free Systems Lab, led by Andy Hall, and a participant in BlueDot's Technical AI Safety course. I built the system this runs on and wrote it up in "Training in the Gaps" (https://getuli.tech/uli-paper.pdf, code at https://github.com/owizdom/uli). Measured on a 16 GB M1 Pro: training is bit-identical over 200 steps with and without serving in between, a job killed twice mid-run matched an uninterrupted run in all 304 arrays, and every dataset, checkpoint and adapter is already hashed. It also ran inside a patched Darkbloom provider on a rented M6. Earlier safety work: Pester, a paper on coding agents that probe their own monitor when falsely blocked (https://github.com/owizdom/pester). Before running the attacks I will have the attack list reviewed by a researcher working on training verification, and I will name the reviewer here once confirmed.
The most likely failure is that training is not bit-identical across chip generations. That is still a result: it rules out plain re-execution for these networks, and I would then measure how large the gap is and whether a tolerance-based check survives the same attacks. The second risk is time, since I am finishing a degree; the minimum tier is scoped to finish regardless.
None from outside funders. I have paid for this project myself since I started it in late October 2026, including the rented Mac used for the Darkbloom run. One application is pending: a BlueDot Rapid Grant for $5,050, which overlaps the Mac rental, the Claude subscription and the laptop in this budget. If it is approved, I will lower this goal by the overlap and post an update here.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!