From Face Recognition To Age Estimation, Potential Biometric Technologies Using ChatGPT-4

3 main points
✔️ Examines the application of large-scale language models to biometric tasks.
✔️ Evaluates the performance of GPT-4 for face recognition, gender detection, and age estimation, and proposes an approach to avoid privacy protection.
✔️ Although GPT-4 shows high performance, it points out the risk of misrecognition and leakage of sensitive information, and suggests the need for further robustness research.

ChatGPT and biometrics: an assessment of face recognition, gender detection, and age estimation capabilities
written by Ahmad Hassanpour, Yasamin Kowsari, Hatef Otroshi Shahreza, Bian Yang, Sebastien Marcel
(Submitted on 5 Mar 2024)
Comments: Published on arxiv.
Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)

code：

The images used in this article are from the paper, the introductory slides, or were created based on them.

Summary

Among the latest technological advances, the emergence of large-scale language models has received a great deal of attention.Through models such asChatGPT, these evolving AIshave shown potential for applications in a wide range of fields such as medicine, education, and even design. ChatGPT, in particular, is known for its versatility and high performance, and is being used in ways we never imagined, from text summarization to image generation. In healthcare, it contributes to patient summarization and early disease prediction, and in education, it enables the creation of unique learning materials. It has also shown its usefulness in creating visuals for design and medical visualization.

However, ChatGPT's capabilities have not yet been fully explored in the area of biometrics. This paper focuses on ChatGPT's potential capabilities in that biometric-related task. In response to privacy concerns such as personal information, ChatGPT has safeguards that avoid direct answers, but this paper proposes a new approach that circumvents this limitation and enables analyses such as facial feature recognition and age estimation.

Through this process, we aim to further improve the accuracy of ChatGPT's interpretation capabilities. Following the initial response analysis, we aim for greater accuracy by using sentiment analysis to evaluate feedback and capture the subtle nuances in the AI's responses. The process is outlined in the figure below. This iterative analysis process demonstrates ChatGPT's superior performance in the biometric task and suggests the potential role that large-scale language models can play in this domain.

CHATGPT for biometric authentication

This paper explores new possibilities for biometric technology by utilizing GPT-4. Starting with face detection, the experiments cover a wide range of applicationsfrom gender detection, age estimation, and face recognition. The figure below shows an example of GPT-4 detecting and counting faces.

To circumvent GPT-4's privacy protection mechanisms that avoid answering direct questions, this paper adopts an approach that elicits deeper analysis by explicitly stating that the images are "generated by AI." Through this ingenious approach, we propose a new method for determining whether each image belongs to the same person.

For each task (face recognition, gender detection, and age estimation), special prompts are set up to bypass GPT-4 safeguards and obtain more precise answers, which are then fed back to GPT-4 itself for further analysis. Through this process, the quality and accuracy of the information provided by GPT-4 is scrutinized and its performance in biometric tasks is thoroughly evaluated.

Facial recognition (the act of)

This paper evaluates the ability of GPT-4 using three datasets, LFW, AgeDB, and CFP-FP, which provide a wide variety of face images. These datasets are ideal for validating GPT-4's ability to accurately identify faces and distinguish between real and AI-generated human faces. The table below shows the accuracy of GPT-4 as evaluated against these three benchmark datasets.

Although GPT-4 is not trained for face recognition, the results shown in this table indicate that it can achieve performance comparable to MobileFaceNet, a pre-trained face recognition model for face recognition. This demonstrates how AI can cope with complex biometric tasks.

Note that the figure below (left) shows a pair correctly identified by GPT-4 as a matching pair, and the figure below (right ) shows two non-matching pairs that were incorrectly classified by GPT-4 as matching pairs. As these samples show, GPT-4 provides an explanation for each prompt. While this may be useful for explainability studies in face recognition, it may also produce misleading output in the case of false positives.

Gender detection

GPT-4 also demonstrated excellent performance in gender detection: in a rigorous evaluation conducted using a dataset published on Kaggle, GPT-4 achieved a 100% accuracy rate. This is based on a balanced dataset containing 2,700 male and 2,700 female images across the entire age range.Using the same dataset, the DeepFace model evaluated performance at 99%. The figure below shows two examples where the DeepFace model failed to classify correctly, but GPT-4 correctly identified them.

In addition, in testing gender detection on synthetic faces generated using Eyes-2-Face technology, GPT-4 challenges the limits of gender classification algorithms by analyzing complex images with a mixture of male and female features.The example shown in the figure below demonstrates this by using a synthetic face created by mixing features from male and female eye regions.

This example may be useful for challenging algorithms with idiosyncratic facial features that do not match standard gender identifiers. In this paper, tests are performed on a set of 200 synthetic faces and show results similar to those obtained using the DeepFace algorithm.

Age Estimation

To evaluate its ability to estimate age, the GPT-4 was tested on 400 real face images using the UTKFace dataset. The evaluation used images representing a variety of age groups. For each image, GPT-4 is asked to estimate the subject's age range. The evaluation criteria are simple: if the actual age of the subject falls within GPT-4's estimated range, it is considered accurate; otherwise, it is considered inaccurate.As a result, out of 400 images, GPT-4 correctly classified 299 images, with an accuracy rate of 74.25%. The figure below shows an example of accurate classification.

The figure below also shows an example of inaccurate classification.

It is then validated using synthetic faces generated by E2F-GAN. The paper uses a dataset containing 100 specially designed AI-generated faces to evaluate GPT-4's recognition and classification capabilities under controlled conditions.The evaluation results show that GPT-4's performance exceeds expectations in terms of accuracy as well as precision and reliability. In particular, throughout this test phase, GPT-4 has shown no errors in synthetic face processing and classification, and has demonstrated a high level of proficiency in handling AI-generated images.The figure below shows an example of age estimation for a synthetic face.

Summary

This paper examines the applicability of large-scale language models such as ChatGPT to biometric tasks. In particular, we focus on ChatGPT's capabilities in biometrics-related tasks, examining its ability to perform face recognition, gender detection, and age estimation. Since biometrics is considered sensitive information, ChatGPT does not respond to direct prompts, and for this reason this paper creates a prompting strategy that bypasses its protection and evaluates its capabilities for biometric tasks.

GPT-4 has the ability to effectively distinguish between different facial features in face recognition tasks and can accurately describe the characteristics of each face. It also shows high accuracy in gender detection, especially for difficult age groups. In terms of age estimation, it tends to predict an age range rather than a precise value, but it does produce predictions that are close to the actual age, especially for the younger population. These experimental results indicate that GPT-4 shows promising performance as a biometric application and suggest that large-scale language models and underlying models may play an important role in biometrics.

On the other hand, it also mentions some points to be aware of when using the GPT-4. In recognition tasks, plausible explanations can be generated that are convincing even in misrecognition scenarios. Also, biometric information is sensitive information and is designed not to respond to direct prompts. However, it has also been suggested that prompt engineering can make large-scale language models vulnerable and potentially leak sensitive information. Future research should include a detailed investigation of the robustness of large-scale language models.

• 
• 
• 

Categories related to this article

• Large Language Models
• Machine Learning
• Deep Learning
• Image Caption
• ChatGPT
• GPT
• Artificial Intelligence