NYU Researchers Detail Age-Altering AI Model

Researchers at the NYU Tandon School of Engineering have unveiled an AI system capable of altering an individual’s appearance to different ages while preserving their distinct facial biometrics. The team’s paper, recently published on the pre-print server arXiv and presented at the IEEE International Joint Conference on Biometrics (IJCB), introduces a novel approach to age transformation using generative AI models.

Led by Sudipta Banerjee, a research assistant professor in the Computer Science and Engineering Department, the team employed a latent diffusion model to achieve identity-retaining age transformations. This approach overcame a common challenge in this domain—limited training data spanning an individual’s lifespan. Instead of a large dataset, the model was trained on a small set of images of a person, accompanied by captions indicating age categories.

By combining biometric data from one set and age-related information from the other, the AI learned the intricate relationship between images and age. Using a text prompt, the model could then simulate the aging or de-aging process by targeting a specific age. This technique, known as “DreamBooth,” involves controlled noise addition and subtraction while respecting the underlying data distribution.

The team conducted comprehensive tests comparing their approach to existing methods. Volunteers were asked to match generated images with actual images of the same individuals, as well as evaluate the model’s performance against the ArcFace facial recognition algorithm. Impressively, the team’s method demonstrated up to a 44 percent reduction in incorrect rejections compared to other techniques.

ArcFace, a loss function in facial recognition introduced by Deng et al. in 2018, played a key role in benchmarking the AI system. Leveraging ArcFace, the researchers optimized their model’s ability to recognize individual facial features and minimize errors.

As the technology landscape continues to evolve, the NYU Tandon research showcases how AI-driven age transformation can have far-reaching implications, from entertainment to security applications. The innovation not only opens new doors for personalized content creation but also underscores the growing capabilities of AI in biometric image generation and deepfakes, which pose an increasing security threat to authentication systems.

Source: Tech Xplore

–

August 28, 2023 – by the FindBiometrics Editorial Team