Abstract
Reactive browser security measures are bypassed by deceptive web pages deployed on unrated domains that employ psychological manipulation vectors. To address this, text and spatial bounding boxes are extracted from a document object model via an asynchronous browser pipeline to generate a layout graph. The layout graph is processed client-side via a local language model using a spatial-textual transformer architecture. Scores for specific persuasion techniques are evaluated via distinct multi-task classification heads. Attention weights are extracted and linked to structural node identifiers to formulate grounded user warnings. Predictive threat detection is achieved based on semantic intent rather than historical domain data.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Yakar, Tamar and Labzovsky, Ilia, "Browser Framework for Deceptive Content Detection via Large Language Model Page Intent Analysis", Technical Disclosure Commons, (July 27, 2026)
https://www.tdcommons.org/dpubs_series/11147