With deepfakes becoming ever-more sophisticated, it’s paramount to adapt existing models for universal detection. This paper demonstrates how adapting the CLIP model using Prompt Tuning greatly improves deepfake detection across diverse datasets.

• VLMs, when adapted properly, can excel in general deepfake recognition.
• Maintaining the textual component of CLIP is critical for performance.
• The adaptation strategy employed here surpasses the SOTA approach.

The work is essential as it shows the adaptability of VLMs like CLIP to combat deepfakes, a growing threat to information integrity. Read the full paper