Skip to main content
Back to timeline
arXivSource publication:

Scam video detection moves past risk cues: MechScam gathers 6,614 videos and ScamCourt's prosecution-defense debate cuts false positives

Synopsis

Online scam videos share reward and payment cues with ordinary promotions and tutorials, so binary labels say little about how a scheme works. MechScam collects 6,614 videos across six scam mechanisms and six non-scam categories, with video-grounded explanations, contextualized risk elements, and misleading arguments built from genuine evidence. On top of it, ScamCourt has trained prosecution and defense debaters argue opposing sides, and a trained judge rules after text quotations and frame references are checked; experiments show it beats larger multimodal models and same-backbone fine-tuned baselines, with fewer false positives and stronger resistance to misleading arguments.

Source-provided article image: Beyond Risk Cues: A Mechanism-Level Benchmark and Verifiable Debate for Scam Video Detection on the Web
arXiv

Interpretation

The evaluation target for scam video detection widens from a single binary label to attributing the scam mechanism and separating scams from benign look-alikes. Earlier datasets supported multimodal scam detection, some with policy-based explanations, but mechanism-level annotations spanning scam content and benign look-alikes were limited. MechScam holds 6,614 videos, 3,079 from VidScam and 3,535 from YouTube_Scam, all re-annotated under one framework, comprising 2,184 scam and 4,430 non-scam videos.

Each video's annotation ties the promised benefit, the action requested of the viewer, and the claimed source of returns together, while keeping qualifying context such as warnings, quotations, and demonstrations. Isolated mentions of money, brands, or external links do not suffice as risk elements; the presence of a risk cue and its interpretation must be judged separately. GPT-5.5 drafts the annotations and humans verify every draft, followed by automatic consistency and citation checks, independent human review, and consensus correction; each video carries a title, description, speech transcript, and up to 64 uniformly sampled frames, 418,442 frames in total.

ScamCourt splits judgment into prosecution opening, defense opening, mutual rebuttal, and a judge's ruling, with the prosecution arguing Scam and a scam category and the defense arguing Non-scam and a benign category. Judges in multi-agent debate must resolve competing interpretations of the same evidence, and genuine material can still be cited to support a misleading conclusion. Before judgment each citation is checked: a text citation must be a contiguous excerpt of the identified source and a frame citation must point to a supplied frame, with all others marked unverified; the judge treats the transcript as advocacy and returns a label, category, content role, and explanation against the original video evidence.

Debaters and judge are trained separately: the debater uses a shared role-conditioned LoRA adapter to argue either side, and the judge learns to rule on cross-fitted debates. A judge trained only on teacher arguments meets a different argument distribution from the trained debater at deployment, and generating judge-training debates with a debater fitted on the same videos adds in-sample effects. Two-fold cross-fitting within the training set is stratified by primary category and source, so each fold's debates come from a debater trained on the complementary fold, and the test-time debater is then trained on the full training set; judge supervision mixes teacher debates, cross-fitted debates, and transcript-free examples for a stratified 25% of training videos.

Perspective

The demonstrated scope is 6,614 YouTube-sourced videos spanning six scam mechanisms and six non-scam categories, with evaluation on binary detection, 12-way fine-grained classification, and resistance to misleading arguments; the dataset and code are released in a public repository.

The material reports that current multimodal large models struggle to separate scam videos from similar non-scam content and remain susceptible to misleading arguments grounded in genuine evidence; these findings come from the constructed benchmark and debate setting, and the material gives no cross-platform or cross-language validation.