Idiap highlights biometrics research, open-source contributions for 2024

Idiap Research Institute has released its 2024 Scientific Report. Its research covers a wide range of digitally relevant topics, several of which are directly applicable to  biometrics and digital ID providers. The research project (RP) on Sustainable and Resilient Societies is exploring ChatGPT for biometrics and how synthetic datasets can expose real identities, while the lab’s more AI-focused RPs look at speech detection technology and bias mitigation in facial recognition.

Explores promising application of LLMs for biometric tasks

The lab’s experiments with ChatGPT examine its capabilities in performing biometric-related tasks, “with an emphasis on face recognition, gender detection, and age estimation.”

For the evaluation, the lab was able to bypass the AI’s safeguard against answering direct prompts about sensitive data. Results have shown that “ChatGPT recognizes facial identities and differentiates between two facial images with considerable accuracy. Additionally, experimental results demonstrate remarkable performance in gender detection and reasonable accuracy for the age estimation tasks.”

In short, ChatGPT can perform facial recognition and associated biometric tasks pretty well.

Study looks at whether synthetic datasets leak real biometrics

For Idiap’s study on synthetic datasets, the lab designed “a simple yet effective membership inference attack to systematically study if any of the existing synthetic face recognition datasets leak any information from the real data used to train the generator model.”

In other words, it looks at whether algorithms used to generate synthetic faces carry over any of the real biometric data they’re trained on.

Idiap says the study “demonstrates privacy pitfalls in synthetic face recognition datasets and paves the way for future studies on generating responsible synthetic face datasets.”

Fair distribution of match scores can help address bias in FRT

In looking at demographic bias in deep learning-based facial recognition systems, Idiap proposes a way to “regularize the training of the face recognition model for demographic fairness, by imposing constraints on the distributions of matching scores.”

The system, it says, respects a pre-defined distribution and penalizes a misalignment of distributions across groups. “The method improves fairness of face recognition models without compromising the recognition accuracy, and does not require extra resources during inference.”

Fairness is a recurring theme across the RPs, which also looks at how to maintain cultural diversity in large language models, and how fine-tuning foundation models without sufficient data increases demographic bias.

AI literacy part of a new approach to learning

The bulk of the research, however, is aligned against the emerging landscape of artificial intelligence. One fascinating study addresses AI literacy and lays out a scheme for twelve “defining competencies.”

“AI literacy has emerged as an important concept, not only for people specialized in technology but also for the public” it says. “Understanding generative AI involves the technical aspects of the algorithms and tools, as well as contextual, ethical, and legal considerations, given the multidisciplinary potential uses and implications of these technologies. We introduced a competency-based model for GenAI literacy, defined by twelve competencies ranging from foundational AI literacy to prompt engineering and programming skills, and to ethical and legal considerations.”

Article Topics

AI  |  biometrics  |  biometrics research  |  facial recognition  |  foundation models  |  Idiap  |  synthetic data