Sheng-Fang Yang, Cheng–Te Li · ACM SIGSAC Conference on Computer and Communications Security (CCS) 2026 · 2026
DOI: 10.1145/3830454.3846418
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Machine-learning screening can prioritize scarce food-factory inspections, yet a convenient feedback rule can convert a transient prediction mistake into persistent training-data corruption: factories predicted as non-illicit are stored as negative examples without independent verification. This promotion of model output to trusted label violates a security-critical provenance boundary, even without an attacker gaining training access. We audit the failure through paired factory-level replays on an inspection-linked chemical-declaration cohort. Matched update policies and representative classifiers separate ordinary updating effects, wrong-label harm, and the distinct impact of model-selected errors on subsequent inspection ranking. Empirical results show that false clearances routinely contaminate retraining and that errors selected from the model's own blind spots are more damaging than comparable random corruption. Thus, high apparent pseudo-label accuracy does not imply safe feedback. A non-flag is an action decision, not verified evidence; its provenance must survive retraining.
arXiv (Cornell University) 2026 · 0 citations
arXiv (Cornell University) 2026 · 0 citations
No comments yet — start the discussion below.