Presentation attacks in the era of cybersecure AI
This lecture explores the rapidly evolving field of biometric security and the increasing reliance on Deep Learning, particularly Convolutional Neural Networks (CNNs), for liveness detection. Biometrics, including fingerprint recognition, have become integral to securing digital and physical access. However, the rise of sophisticated presentation attacks, which exploit vulnerabilities in biometric systems, poses significant challenges. The lecture delves into how modern Fingerprint Presentation Attack Detection (FPAD) systems leverage Machine Learning and Deep Learning to effectively counteract artificial fingerprint replicas.
Despite these advancements, adversarial attacks remain a pressing concern. These attacks, designed to deceive detectors by manipulating input data, highlight critical vulnerabilities in CNN-based systems. The discussion covers how adversarial presentation attacks, once deemed unrealistic due to the need for access to communication channels between sensors and detectors, have now become feasible. Recent studies have demonstrated the transfer of fingerprint adversarial attacks from digital to physical domains, revealing a new realm of security threats.
The lecture introduces novel procedures aimed at enhancing the robustness of physical adversarial presentation attacks using explainability techniques. These procedures make attacks more adaptable across different fingerprint scanners and adversarial algorithms, and viable in black-box scenarios. By assessing the performance impact on state-of-the-art PAD modules and integrated Automated Fingerprint Identification Systems (AFISs), the lecture illustrates the feasibility and implications of these threats, not only from a technical point of view, but as from a legal as well as from an ethical perspective.