Face Liveness Detection under Processed Image Attacks

Abstract

Face recognition is a mature and reliable technology for identifying people. Due
to high-definition cameras and supporting devices, it is considered the fastest and
the least intrusive biometric recognition modality. Nevertheless, effective spoofing
attempts on face recognition systems were found to be possible. As a result, various anti-spoofing algorithms were developed to counteract these attacks. They are
commonly referred in the literature a liveness detection tests. In this research we highlight the effectiveness of some simple, direct spoofing attacks, and test one of
the current robust liveness detection algorithms, i.e. the logistic regression based face liveness detection from a single image, proposed by the Tan et al. in 2010, against malicious attacks using processed imposter images. In particular, we study experimentally the effect of common image processing operations such as sharpening and smoothing, as well as corruption with salt and pepper noise, on the face liveness detection algorithm, and we find that it is especially vulnerable against spoofing attempts using processed imposter images. We design and present a new facial database, the Durham Face Database, which is the first, to the best of our knowledge, to have client, imposter as well as processed imposter images. Finally, we evaluate our claim on the effectiveness of proposed imposter image attacks using transfer learning on Convolutional Neural Networks. We verify that such attacks are more difficult to detect even when using high-end, expensive machine learning techniques.