Morphing Resilient Face Recognition by Informed Frequency Selection

Face recognition (FR) systems have established themselves as a reliable system for verifying identities, especially in security scenarios such as at border crossings. However, existing systems have been proven to be vulnerable to morphing attacks, which refer to manipulated face images that combine the facial features of two distinct individuals in one image. These morphed images can be matched by automatic systems or humans to both individuals, creating a major security risk. In this work, we propose a training- and data-free ad-hoc approach, MR-FR, to enhance the FR resilience to such attacks. This leverages recent explainability insights into the behavioral differences of FR systems when processing morphing attack images. By informed frequency-based manipulation of potential morphing attack images, we were able to reduce the FR vulnerability in terms of MMPMR up to 55.1 percentage points, while having only a minor accuracy trade-off.