Linguistic Profiling of Deepfakes: An Open Database for Next-Generation Deepfake Detection

TL;DR

This work addresses the growing challenge of detecting and explaining deepfakes produced by text-to-image models. It presents DFLIP-3K, an open resource with roughly 300K deepfake samples from ~3K generative models and about 190K linguistic prompts, designed to support three tasks: deepfake detection, model identification, and prompt prediction. Through a Flamingo-based benchmark and extensive experiments, the authors show that vision-language models outperform vanilla vision approaches and can provide trustworthy, interpretable evidence via reconstructed prompts and model signals. The dataset and benchmark aim to foster transparent, explainable deepfake detection in the face of rapidly evolving generative technologies, while acknowledging biases, misuse risks, and the need for responsible dissemination.

Abstract

The emergence of text-to-image generative models has revolutionized the field of deepfakes, enabling the creation of realistic and convincing visual content directly from textual descriptions. However, this advancement presents considerably greater challenges in detecting the authenticity of such content. Existing deepfake detection datasets and methods often fall short in effectively capturing the extensive range of emerging deepfakes and offering satisfactory explanatory information for detection. To address the significant issue, this paper introduces a deepfake database (DFLIP-3K) for the development of convincing and explainable deepfake detection. It encompasses about 300K diverse deepfake samples from approximately 3K generative models, which boasts the largest number of deepfake models in the literature. Moreover, it collects around 190K linguistic footprints of these deepfakes. The two distinguished features enable DFLIP-3K to develop a benchmark that promotes progress in linguistic profiling of deepfakes, which includes three sub-tasks namely deepfake detection, model identification, and prompt prediction. The deepfake model and prompt are two essential components of each deepfake, and thus dissecting them linguistically allows for an invaluable exploration of trustworthy and interpretable evidence in deepfake detection, which we believe is the key for the next-generation deepfake detection. Furthermore, DFLIP-3K is envisioned as an open database that fosters transparency and encourages collaborative efforts to further enhance its growth. Our extensive experiments on the developed benchmark verify that our DFLIP-3K database is capable of serving as a standardized resource for evaluating and comparing linguistic-based deepfake detection, identification, and prompt prediction techniques.

Figures (20)

• Figure 1: DFLIP-3K enables linguistic profiling of deepfakes, including the assessment of authenticity (deepfake detection), the identification of source model (deepfake identification), and the prediction of source prompt (prompt prediction).
• Figure 2: A comparison of DFLIP-3K against the early and recent deepfake datasets. It collects deepfakes from about 3K generative models, representing the largest scale in terms of the number of deepfake models.
• Figure 3: Statistics of DFLIP-3K that consists of six big families of generative models. Personalized Diffusions include deepfakes from 3,434 known generative models and a set of unknown models.
• Figure 4: The proposed benchmark setting on DFLIP-3K. Digits outside parentheses are for deepfake detection, and the ones in parentheses are for the other two tasks. Personalized Diffusion (PD) is divided into PD1 (51 Models) for in-distribution training/test, and PD2 ($\approx$ 3K models) for out-of-distribution test.
• Figure 5: DFLIP-3K examples generated by the six groups of generative models. More examples are presented in the Supplementary Material.