Deepfake detection and attribution systems that rely only on visual features tend to generalize poorly to new generative methods, the paper argues. The authors trace this weakness to an exclusive focus on visual information, which leaves models unprepared for generators they have not encountered before.
GazeCLIP addresses this by introducing gaze as an additional cue. The model builds on CLIP, the vision-language architecture, and uses adaptive-enhanced fine-grained language prompts to tie visual content, gaze behavior, and textual descriptions together. The aim is to make both attribution—identifying which generator produced a fake—and detection—deciding whether media is fake—more robust.
The source abstract is truncated, so specific experimental results and comparisons with prior methods are not available in the text provided. This summary covers only the motivation and architecture described in the available portion.