AEyeDE: An Attention-Based Attribution Framework for AI-Generated Text Detection

In three linesAEyeDE introduces an attention-based attribution framework for detecting AI-generated text using attention matrices from a proxy Transformer model. A lightweight CNN learns discriminative representations from these attribution maps. The method outperforms text-only baselines, shows strong generator-specific detection, and demonstrates robustness under cross-dataset transfer and spelling perturbations.

Summary generated by Claude — human-verified