Large language models, or LLMs, have shown poorer performance in deepfake recognition than those demonstrated by state-of-the-art detection algorithms, but their natural language processing may actually make them more effective in the future, according to a study led by the University at Buffalo, in collaboration with the University at Albany and the Chinese University of Hong Kong, presented last week at the IEEE/CVF Conference on Computer Vision & Pattern Recognition. Most people think of artificial intelligence in association with ChatGPT and deepfakes. AI-generated text and images dominate social media feeds and websites and are often used to spread unreliable and misleading information. In the study, the researchers asked LLMs, including OpenAI’s ChatGPT and Google’s Gemini, to spot deepfakes of human faces. “What distinguishes LLMs from existing detection methods is the ability to explain their findings in a way that humans understand, such as identifying the wrong shadow or a mismatched pair of earrings,” Siwei Lyu said , of the Department of Computer Science and Engineering, within the UB School of Engineering and Applied Sciences and lead author of the study.
“LLMs were not designed or trained for deepfake detection, but their semantic knowledge makes them well-suited for this purpose, so we expect to see more efforts toward this application,” Lyu continued. The latest versions of ChatGPT and other LLMs can also analyze images. These multimodal LLMs use large databases of photos with captions to find relationships between words and images. “Humans do this too: whether it’s a stop sign or a viral meme, we’re constantly assigning semantic descriptions to images,” said Shan Jai, a UB Media Forensic Lab scientist and the study’s first author. “This way, the images become their language“, Jai continued.The Media Forensics Lab team decided to test whether GPT-4 with vision, GPT-4V, and Gemini 1.0 could distinguish between real faces and AI-generated faces. The scientists gave them thousands of images of real and deepfake faces and asked them to identify any signs of manipulation or synthetic artifacts. ChatGPT was accurate 79.5% of the time in detecting synthetic artifacts in images generated by latent diffusion and 77.2% of the time in images generated by StyleGAN.
“These are comparable results to previous deepfake detection methods, so with timely and proper guidance, ChatGPT can do a pretty decent job of detecting AI-generated images,” said Lyu, who is also co-director of the UB Center for Information Integrity. Furthermore, ChatGPT is able to explain its decisions in simple language. When given an AI-generated photo of a man wearing glasses, the model correctly detected that “the hair on the left side of the image is slightly blurred” and “the transition between the person and the background is a little blunt and lacks depth. “Existing deepfake detection models tell us the probability that an image is real or fake, but very rarely tell us why they came to this conclusion,” Lyu said.
“And – added Lyu – even if we analyze the mechanisms underlying the model, there will be features that we simply will not be able to understand.” “Meanwhile, everything ChatGPT produces is understandable to humans,” Lyu highlighted. This is because ChatGPT bases its analysis only on semantic knowledge. While traditional deepfake detection algorithms distinguish real from fake by training on large sets of images labeled real or fake, LLMs’ natural language capabilities give them a kind of commonsense understanding of reality, including symmetry typical of human faces and the appearance of real photographs. “Once the visual component of ChatGPT understands that an image is a human face, the linguistic component can infer that a face typically has two eyes, and so on. The linguistic component provides a deeper connection between visual and verbal concepts“, Lyu pointed out. The study claims that ChatGPT’s semantic knowledge and natural language processing make it an easier to use deepfake tool for both users and developers. “Usually, we take detection insights deepfakes and convert them into programming language; now, all this knowledge is present within a single model and you just need to use natural language to bring it out,” Lyu noted. “ChatGPT’s performance was well below the latest deepfake detection algorithms, which have accuracy rates in the 90s to 30s,” Lyu said. This is partly because LLMs are unable to capture statistical differences at the signal level, which are invisible to the human eye but are often used by detection algorithms to locate images generated by artificial intelligence.”ChatGPT focused only on anomalies at the semantic level“, Lyu explained. “In this way, the semantic intuitiveness of ChatGPT results could be a double-edged sword for deepfake detection,” Lyu added. And other LLMs may not be as good at explaining their analyses.
While it achieved comparable results to ChatGPT in identifying the presence of synthetic artifacts, the evidence supporting Gemini was often nonsensical, such as indicating non-existent moles. Another drawback is that LLMs have often refused to analyze the images. When asked directly whether a photo was generated by AI, ChatGPT typically responded with: “Sorry, I can’t respond to that request.” “The model is programmed not to respond when it does not reach a certain level of confidence.“, Lyu commented. “We know that ChatGPT has information relevant to deepfake detection, but again, a human operator is needed to stimulate this part of its knowledge base,” Lyu noted. “Prompt engineering is effective, but not very efficient, so the next step is to go down a level and fine-tune LLMs for this specific task,” Lyu concluded.