Can ChatGPT Detect DeepFakes? A Study of Using Multimodal Large Language Models for Media Forensics

View on arXiv ← Back to list

Authors: Shan Jia, Reilin Lyu, Kangran Zhao, Yize Chen, Zhiyuan Yan, Yan Ju, Chuanbo Hu, Xin Li, Baoyuan Wu, Siwei Lyu

Published: 2024-03-21 01:57:30+00:00

AI Summary

This research explores the feasibility of using multimodal large language models (LLMs) for deepfake detection, specifically focusing on AI-generated images. The study shows that LLMs, despite not being designed for this task, can effectively identify AI-generated images through prompt engineering and achieve a reasonable detection accuracy.

Abstract

DeepFakes, which refer to AI-generated media content, have become an increasing concern due to their use as a means for disinformation. Detecting DeepFakes is currently solved with programmed machine learning algorithms. In this work, we investigate the capabilities of multimodal large language models (LLMs) in DeepFake detection. We conducted qualitative and quantitative experiments to demonstrate multimodal LLMs and show that they can expose AI-generated images through careful experimental design and prompt engineering. This is interesting, considering that LLMs are not inherently tailored for media forensic tasks, and the process does not require programming. We discuss the limitations of multimodal LLMs for these tasks and suggest possible improvements.

Key findings

Multimodal LLMs demonstrated a capability to distinguish between real and AI-generated images, achieving AUC scores comparable to some early deepfake detection methods. However, their accuracy in recognizing genuine images was lower, and performance was highly dependent on prompt engineering. The LLM's reliance on semantic understanding rather than signal-level analysis offers a more human-interpretable approach but limits its competitive edge against state-of-the-art methods.

Approach

The researchers used multimodal LLMs (like GPT-4V and Gemini) to classify images as real or AI-generated. They experimented with various prompts to determine which were most effective at eliciting accurate classifications and justifications from the LLMs. Performance was evaluated using metrics like AUC and accuracy.

Datasets

FFHQ dataset (real faces) and DF-3 dataset (AI-generated faces, including StyleGAN2 and Latent Diffusion models, with and without post-processing).

Model(s)

OpenAI's GPT-4V Vision model and Google Gemini 1.0 Pro API.

Author countries

USA, China

← Previous