Alexander Nguyen, Dylan Gresham, Gaby G. Dagher and Min Long submitted SciCheck to arXiv on 6 November 2026. The framework proposes checking deep learning training through recorded training plans, gradient statistics and sampled re-execution. The authors frame the problem as scientific verification: training configurations evolve during experimentation, and honest runs produce different weights, complicating checks based on recomputation.[1]

Scientists post a blueprint for each candidate configuration to a provenance ledger. A verifying scientist then uses two checks: a statistical judge compares per-step gradient statistics with the verifier’s reference run, while a spot check re-executes randomly selected training segments. Checkpoints are committed to the ledger one epoch at a time.[1]

The abstract reports experiments on an MLP, a CNN and an autoencoder with 100,000–500,000 parameters, testing gradient sign flipping, label poisoning and fabricated runs. It reports F1 scores of 0.966, 0.906 and 0.976, with 0–4.4% of honest runs rejected. These are author-reported results. The authors say spot checks account for most accuracy against the tested attacks; fabricated runs replaying genuine statistics appeared no more anomalous than honest runs to the statistical detectors they tested.[1]

[1]: SciCheck: A Framework for Verifying the Integrity of Deep Learning Training. arXiv, 6 November 2026. https://arxiv.org/abs/2611.09889

Sources

  1. SciCheck: A Framework for Verifying the Integrity of Deep Learning TrainingarXiv · 6 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

b8b4709d93f5d12cb3567f57564807c327987d60e602d011307b788d86497110