Recently, a study conducted by researchers from the Norwegian University of Science and Technology, Mizani, and Idiap Research Institute revealed that GPT-4 possesses capabilities for facial recognition, gender determination, and age estimation, with accuracy levels rivaling those of specialized facial recognition algorithms. Surprisingly, GPT-4 has not been specifically trained to perform these tasks.
Image Source: The image was generated by AI, provided by the image authorization service Midjourney
In this study, researchers tested the biometric capabilities of GPT-4. In the gender recognition test, GPT-4 achieved a perfect accuracy rate of 100% on a balanced dataset of 5400 images, surpassing the DeepFace model, which has an accuracy rate of 99%. This result is truly impressive.
Regarding age estimation, GPT-4's performance is also satisfactory, with a correct identification probability of 74.25%. However, researchers noted that GPT-4 tends to provide broader age ranges for individuals over 60, while its estimations for younger individuals are relatively accurate.
Nonetheless, the study also revealed some security vulnerabilities. Researchers discovered that they could bypass GPT-4's built-in protection mechanisms with a simple method: merely claiming that the input image is AI-generated, GPT-4 would begin analyzing real photos. This finding raises further concerns about the security of large language models.
Although the study shows that GPT-4 performs exceptionally well on biometric tasks, the authors warn that it should not be fully relied upon for identification, as it sometimes provides convincing but incorrect descriptions. In fact, GPT-4's biometric capabilities are not new; OpenAI previously mentioned this in the "Be My Eyes" application, but the facial recognition feature was disabled for security reasons.
The new findings of this study highlight that GPT-4's security guidelines can be bypassed with simple tricks, and its accuracy in biometric tasks is astonishing.
Key Points:
🌟 GPT-4 achieved a perfect accuracy rate of 100% in gender recognition tests, outperforming the DeepFace model.
📊 In age estimation, GPT-4's accuracy is 74.25%, but it may provide broader ranges for older individuals.
🔒 The study found that GPT-4's security mechanisms can be bypassed, necessitating stronger research into the security of large language models.