FOX separates image-detection verdicts from explanations

FOX separates image-detection verdicts from explanations

Authors Nikos Giakoumoglou, Paschalis Giakoumoglou, Tania Stathaki and Symeon Papadopoulos introduce FOX, a system for detecting AI-generated images and producing forensic explanations. They argue that adapting a detector to a detection distribution can make it memorize generator-specific fingerprints, while jointly training classification and language generation can weaken the visual representation used for detection. FOX first builds and locks a DINOv3 detector, then uses its representation to guide a lightweight VLM in generating explanations. The authors’ central design constraint is that explanation training cannot alter the detector’s verdict. They report a mean AUC of 97.4 across 14 unseen generators, with BERTScore results of 69.8 on XPlainVerse and 56.1 on LOKI. These are reported benchmark results; the abstract does not establish performance in real-world use. How do the authors evaluate whether FOX’s explanations accurately reflect the evidence behind its verdicts?

Publication: 666bcd2e-8444-4578-8578-27ec345d07ce

attribution — pass: The reading correctly names the four listed authors and attributes the FOX design, motivation and benchmark figures to them as their own claims, not as established fact.

Authors:Nikos Giakoumoglou, Paschalis Giakoumoglou, Tania Stathaki, Symeon Papadopoulos

summary — uncertain: A non-specialist cannot follow the central terms. DINOv3, VLM, AUC, BERTScore, unseen generators and generator-specific fingerprints are never explained in plain language or flagged for tooltips. The reading also drops the abstract's stated stakes: realistic images, editors that leave no splice boundary, and black-box detectors with unverifiable verdicts. Without these, readers miss why explainable detection matters. Smallest repair: add one sentence on the black-box and editing problem, and gloss VLM, AUC and BERTScore (or add term-library entries).

Most detectors are black boxes with unverifiable verdicts and generalize poorly

date — pass: originalPublished 2026-11-03 matches the v1 submission date on the arXiv page.

[Submitted on 3 Nov 2026]

support — pass: The claims closely track the abstract. Saying joint training can weaken the visual representation is a fair paraphrase of the abstract's point that explainable methods forgo the generalization advantage because the representation is subjected to the language objective. The caveat that benchmarks do not establish real-world performance is appropriate. The question follows from the evidence without being leading.

explainable methods forgo this advantage: jointly training one multimodal model for classification and language generation subjects its visual representation to the language objective

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.