Hummingbird: Human-machine integration for biometric authentication

点击翻译按钮获取中文摘要

We live in a technological age in which we can use our voice as a password to access online banking, and our children can pay for school lunches with a fingerprint. Biometrics, which reflect our physiological or behavioural characteristics, are now common as a way to prove our identity in order to access secure information, services or spaces. Given the important uses associated with biometrics, there is a fundamental need for accuracy in biometric analysis in order to encourage trust amongst both citizens and service providers. The feasibility study undertaken within the HUMMINGBIRD project will provide a human-inspired framework to address both needs.The recent publication of two high profile report converge to make this endeavour timely and necessary. The first is the UK Governmental review on 'Future Identities', which recognised the transformative effect that digital technologies are having on identity. In particular, it noted the myriad of ways we now have to convey our identity, and to have it spoofed. The second is the UK Parliamentary Select Committee review on 'The Current and Future Uses of Biometrics' which highlighted two necessary future steps for biometric analysis: Analysis should draw on behavioural as well as physiological measures; and it should take full advantage of the combination of data across multiple biometrics and across decision makers in order to improve decision-making.To address all factors, we propose an exciting project that will deliver a human-inspired multi-expert, multi-modality framework for biometric analysis. This will satisfy three aims: First, it will deliver enhanced algorithms for automated biometric analysis by incorporating successful strategies used by humans. Second, it will deliver a method of combining decisions made by humans and (enhanced) algorithms in order to boost accuracy. Third, it will deliver the potential to combine multiple biometrics, providing resilience in scenarios in which a single modality may be sub-optimal.The HUMMINGBIRD project team possesses a unique combination of skills to explore this idea and indeed, we build on recently published theoretical work on this topic. In this proposal, we examine two biometrics - face and voice - which reflects the move to combine static physiological measures (facial images) and dynamic behavioural measures (temporal voice samples). We also concentrate on two decision-makers - the human and state-of-the-art automated algorithm - providing direct relevance to scenarios in which the human must be part of the decision-process (such as in forensic decisions). Our work will establish the fundamental performance levels of humans and machine algorithms when recognising faces and voices under optimal and sub-optimal presentation conditions. It will then seek to enhance the machine algorithms through incorporation of human rules and heuristics. Such a move offers the potential to boost accuracy and efficiency by streamlining automated solutions. More importantly, it exploits the fact that humans can outperform machine algorithms under some conditions, such as when trying to recognise a face under dim light, or a voice amidst noise. Finally, our work will apply an innovative data fusion model to combine the decisions of humans and machine algorithms from one biometric, and then from multiple biometrics. This novel and creative element of our work addresses issues of accuracy, disagreement resolution, and resultant confidence in an identity decision, when the situation is inherently uncertain. Arguably, biometrics reflect identity more directly than token or password systems because they rely on who we are rather than what we have or know. As such, biometric analysis is likely to remain a mainstay of identity management. The HUMMINGBIRD project presents real promise as a way to improve accuracy and confidence in that analysis, enabling accuracy of, and trust in, identity management as technology advances.