ChatGPT can recognize ‘facial identities,’ perform age estimation: research
The large language model (LLM) ChatGPT “recognizes facial identities and differentiates between two facial images with considerable accuracy,” according to a newly released study exploring the potential in applying LLMs to “biometric tasks.”
The paper, “Chatgpt and Biometrics: an Assessment of Face Recognition, Gender Detection, and Age Estimation Capabilities” delivers on its title. “While the applications of LLMs, such as ChatGPT, have been studied for different tasks, their capabilities for biometrics have not been explored,” it reads.
The research team of academics from Idiap Research Institute, Mizani Research Institute and the Norwegian University of Science and Technology (NTNU) says ChatGPT “avoids answering questions regarding sensitive information (e.g., privacy-related) because of safeguards. Along the same lines and since biometric data are considered as privacy-sensitive information, ChatGPT does not provide direct answers to prompts regarding biometric information.”
But what about indirect queries? Using “a crafted prompt designed to bypass the safeguard mechanisms of ChatGPT,” the researchers were able to test the program’s biometric capabilities – which they found to be significant.
“Our study reveals that ChatGPT recognizes facial identities and differentiates between two facial images with considerable accuracy,” says their summary. “Additionally, experimental results demonstrate remarkable performance in gender detection and reasonable accuracy for the age estimation tasks.”
The team trained ChaGPT-4 on datasets consisting of thousands of images of both real and synthetic faces and tested its biometric capabilities. The paper details their methods for face recognition, age estimation and gender detection. But it is not until the conclusion that it states explicitly the implications inherent in the experiment – that safeguards on ChatGPT’s biometric functions are relatively easy to demolish.
“Our experimental results reveal the effectiveness of ChatGPT in recognizing facial identities and differentiating them, as well as promise in gender detection and age estimation,” the report says. “We crafted a prompting strategy to bypass its safeguards and assess its capabilities for biometrics tasks. Our study, however, shows that by prompt engineering, LLMs may become vulnerable and disclose sensitive information, and sheds light on the demand for further research on the robustness of LLMs.”
Facial recognition success with GPT-4
ChatGPT is likely being prodded at will in labs, offices and dorm rooms around the world, as users look to understand the full range of its abilities. In The Conversation, Robin Kramer, senior lecturer in the School of Psychology, University of Lincoln, describes his own testing on GPT-4 and its ability to perform various biometric tasks.
Using the “reading the mind in the eyes” test (developed by the Autism Research Centre at the University of Cambridge), the “Glasgow face matching test,” and facial recognition, Kramer discovers that ChatGPT “seems well-equipped to perform tasks related to the recognition and identification of human faces – including their expressions. It performed as well as or even better than people do for these three tests, at least.”
Article Topics
accuracy | age estimation | biometric identification | biometrics | biometrics research | ChatGPT | facial recognition | gender recognition | Idiap
Comments