A Multimodal Detection Model for Online Fraud Targeting Adolescents

Online deception targeting adolescents increasingly manifests through fabricated reward schemes, deceptive hyperlinks, identity impersonation, virtual item trading scams, suspicious authentication pages, and social invitation manipulation. Detecting such threats is challenging because deceptive intent is often distributed across textual content, URL structures, visual webpage cues, and interaction metadata. Existing detection methods primarily address general adult-oriented scenarios, while single-modality or late-fusion models have limited capacity to capture cross-modal inconsistencies and provide reviewable decisions for adolescent safety applications. To address these limitations, this study proposes a multimodal detection framework integrating text, URL, screenshot, and lightweight metadata features. The architecture incorporates a youth-scenario risk encoder, a cross-modal risk-consistency attention module, and an uncertainty-aware decision layer. Experiments were conducted using publicly available deception-related datasets and a manually annotated adolescent-risk subset. The proposed model achieved an accuracy of 0.905 ± 0.007, a macro-F1 of 0.901 ± 0.008, and an AUROC of 0.943 ± 0.006. Compared with Text-BERT, the macro-F1 improved by 3.3 percentage points, and compared with late-fusion multimodal detection, it improved by 1.8 percentage points. The false-positive rate was reduced to 0.071 ± 0.006. These results indicate that cross-modal consistency modeling can enhance detection reliability, interpretability, and privacy-conscious decision support for adolescent online safety systems.