AdaptPrompt with Diffusion Set: A Unified Framework for Generalizable Deepfake Detection

Abstract

Deepfake detection focuses on identifying synthetic media. It has critical applications in cybersecurity, misinformation mitigation, digital forensics, and media authentication. Re- cent developments in deepfake detection have achieved impressive performance, leveraging deep learning models to distinguish between real and synthetic content. However, recent developments in diffusion-based generative models and publicly available tools such as Stable Diffusion and DALL-E pose immediate challenges to existing detection techniques. Diffusion models generate photorealistic and high-resolution content with less observable or detectable artifacts and are, therefore, more difficult to detect using traditional deepfake detection techniques. In this thesis, we present a comprehensive study of existing deepfake detection tech- niques and adapting large vision-language models, i.e., CLIP, to generalizable deepfake detection. We introduce the Diffusion Set, a new dataset of 100k diffusion-generated fake images and 100k real images. Our experiments reveal that detectors trained on Diffusion Set outperform detectors trained on GAN-based datasets. To further enhance deepfake detection, we introduce a new transfer learning strategy that learns randomly initialized prompts and a lightweight adapter network while having CLIP frozen. Extensive experi- ments confirm its efficiency, and we also investigate the impact of dropping specific CLIP layers on detection accuracy. Our study utilizes the Diffusion Set for training and evaluates models on 25 unseen test sets, covering images synthesized by GAN-based models, diffusion-based models, and commercially available tools. Beyond large-scale training, we assess model performance in few-shot settings, where models are trained with only a small fraction of the dataset (e.g., 320 real and 320 fake images), providing insights into their adaptability under data constraints. Additionally, we extend our analysis beyond classification by exploring image attribu- tion and training models in a few-shot setting to attribute images to specific generators such as BigGAN, StarGAN, and Stable Diffusion. Our findings showcase the robustness of CLIP-based models in deepfake detection and their ability to generalize across unseen gen- erative techniques. We also investigated the feasibility of using the same transfer learning strategy for attribution, with experimental results that demonstrate its high effectiveness in closed set attribution.