Validation ranking hid one detector’s false positives on genuine photos
Source check: Source checked. Not independent scientific confirmation.
Finding: An OpenCLIP detector selected at 98.43% validation accuracy flagged 1,143 of 5,413 genuine photographs in a held-out OpenForensics-derived test set. A checkpoint ranked lower on validation flagged 275 genuine images and missed more manipulated images. The test contained 10,905 images.
Technology: OpenCLIP ViT-L/14; stochastic weight averaging
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Fine-tuned OpenCLIP ViT-L/14 checkpoints classified genuine and manipulated photographs. The trained regular and averaged checkpoints could be compared on the same held-out classification set.
- Test and result
- The validation-selected detector had 98.43% accuracy on validation and 88.85% on a 10,905-image held-out OpenForensics-derived test. Test errors included 1,143 false accusations among 5,413 genuine photographs. The final weight average, lower-ranked on validation, had 94.76% accuracy and 275 false positives but more missed manipulations.
- Limitation / still unknown
- The result comes from a single training run and two disjoint partitions of one derived corpus. The pattern is not independently replicated, and illustrative examples do not establish what caused the false positives.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.37874v1 · https://arxiv.org/html/2610.37874v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: arXiv v1 §§IV-A, IV-C, IV-G–H, Tables III, V and IX, and V-D–G.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.