Performance variation of morphed face image detection algorithms across different datasets

In past years, different researchers have shown the vulnerability of face recognition systems to attacks based on morphed face images. More recently, first morph detection subsystems have been proposed to automatically detect this kind of fraud. While some algorithms have been reported to reveal practical detection performance on individual datasets a systematic analysis of proposed detectors with respect to their robustness across different databases has remained elusive. In this work, we evaluate the performance of different morph detection algorithms across disjoint datasets of 2,745 bona fide and 14,337 automatically generated morphed face images. Within a generic evaluation framework a systematic robustness estimation scheme is proposed to identify reliable detection algorithms. Finally, the robustness of algorithms which have been determined as most promising is verified on another disjoint dataset. Hence, this paper represents the first attempt towards a comprehensive cross-database performance evaluation and a systematic evaluation of the robustness of morphed face image detection algorithms.