Visual Rhythms-based Video Face Spoofing Detection Using Deformable Self-Attention ResNet Model

There are primarily three types of techniques that may compromise face authentication: print, replay, as well as 3D mask. It seems that video replay assaults are the most difficult to detect out of all of them. Even if there are a lot of ways to prevent spoofing assaults, especially with high-quality videos, a sophisticated detector is still required. This study describes video-based spoofing techniques by analyzing the frequency-domain noise residual and extracting discriminative features with a dynamic texture descriptor. We present a potential detector that outperforms the state-of-the-art on the most difficult video spoofing dataset. On the other side, inexperienced camera operators often introduce undesired camera shake or motion into their recordings. The goal of video face spoofing methods is to improve the quality of the captured image by identifying and fixing any imperfections that may have occurred during capture. To assess video face spoofing techniques qualitatively, we offer and examine a new representation using Visual Rhythms (VR). The efficiency of the visual representation as a qualitative metric for analyzing spoofing attacks is shown via experiments done on various video sequences. Using ResNet-18, our database achieved a spoofing detection accuracy of 99.75%. Spoof detection abilities were better for photographs with a pinch angle, close distance images, with replay assaults compared to front images, far distance visuals, with print attacks, respectively, according to the trial findings for different circumstances. Compared to testing with6 other databases (DBs) before training with our DB, the results of the cross-database verification showed an improvement in performance. Since there was little variation in the findings of the cross-device verification based on camera type, we can say that the detection accuracy is unaffected by the kind of image sensor. We also provide a method to objectively determine a measure derived from VR. The experiment found that the Total Error Rate (HTER) for the four independent data sets was 3.5%, whereas the synthesized data set had an HTER of 1%.