Presentation attack detection on face recognition system in mobile devices

Resumo

Face recognition technology is nowmature enough to reach commercial products. There are more and more scenarios where we use this technology (airport border controls, to unlock our mobile device, to register in online services, etc.). Although facial recognition technology for verification has surpassed even humans on several benchmarks, there are still many challenges ahead to become a secure, usable, fair, unbiased and privacy-friendly technology. In this thesis we study one of the most vulnerable parts of the current facial recognition systems, the spoofing attacks and the countermeasures to avoid impersonation attempts. In this regard, face Presentation Attack Detection (face-PAD) is a key component to provide trustable facial access to digital devices. Despite the success of several face-PAD works in publicly available datasets,most of them fail to reach the market, revealing the lack of evaluation frameworks that represent realistic settings. We provide an extensive analysis of the generalization problem in face-PAD, jointly with two proposal for evaluate state-of-the-art algorithms. First of all, we introduce a new evaluation framework that for the first time takes into account crucial parameters in the production deployment of the systems, such as framerates and response times requirements. On the other hand, we propose an evaluation strategy based on the aggregation of most publicly-available datasets and a set of novel protocols to cover the most realistic settings, including a novel demographic bias analysis. Besides, we provide a new fine-grained categorization of presentation attacks and instruments, enabling higher flexibility to assess the generalization of different algorithms under a common framework. As a result, we present the GRAD-GPAD, a comprehensive and modular framework to evaluate the performance of face-PAD approaches in realistic settings, enabling accountability and fair comparison of most face-PAD approaches in the literature. Moreover, all evaluation frameworks are released as open source to encourage reproducible research and facilitate the implementation of fairer face-PAD evaluation strategies.