Kaydedildi:
| Asıl Yazarlar: | , , |
|---|---|
| Materyal Türü: | Recurso digital |
| Dil: | |
| Baskı/Yayın Bilgisi: |
Zenodo
2026
|
| Konular: | |
| Online Erişim: | https://doi.org/10.5281/zenodo.20393375 |
| Etiketler: |
Etiketle
Etiket eklenmemiş, İlk siz ekleyin!
|
İçindekiler:
- Dark patterns are the deceptive interface designs that manipulate users into harmful actions. Even though reg- ulators now impose very large fines, t e chnical d e fenses still remain limited: recent studies show a 54.5% coverage gap, with only 31 of 68 known dark-pattern types being detected by the tools available. This is due to three main reasons. The datasets being offered in public are small and narrow, typically a few thousand examples covering at most 15 to 20 pattern types. Even though detectors analyze static screenshots or isolated text, they are not acquainted with multi-step flows l i ke R o ach Motel cancellations and hidden subscriptions. Meanwhile, advanced Al offers more personalized deception and weakens defenses through adversarial attacks on NLP and vision models. This work surveys the existing text-based, visual, multimodal, and conversational detection methods, comparing approaches, datasets, metrics, and limitations. Building on this analysis, it proposes a four-engine framework that integrates multiple variable like DOM, visual, linguistic, and behavioral signals, and outlines a 10,000-example, multi-platform dataset aligned with a 245-pattern ontology. Finally, it sketches privacy-preserving and adversarially robust deployment strategies for real-time protection.