Optimal feature-level fusion and layered k-support vector machine for spoofing face detection

摘要

The recognition frameworks are highly vulnerable to spoofing attacks and this vulnerability generates an effective security concerned issues in biometric domain. Moreover, some of the earlier proposed approaches have attained attractive results with intra test (i.e. by training and testing the system on same database) evaluation done to detect the face spoofing attack. Consequently, most of these techniques generate incorrect decision on the recognition of genuine faces with unseen attacks in case of inter test evaluation (i.e. the system is trained on one database and then tested on another database). However, this impact is considered as a major difficulty in the highly focused biometric anti-spoofing research domain. In this work, we propose a multimodal biometric framework for the accurate recognition of fake face from genuine face. Initially, face image features which are coupled to the color spaces HSV and YCbCr are extracted with EDGHM-SURF (Enhanced Discrete Gaussian-Hermite Moment based Speed-up Robust Feature) descriptor. Then, a newly developed method of feature-level fusion using OGWO (FLFO) is used to fuse these extracted features. This method utilized the OGWO (Oppositional Gray Wolf Optimization) algorithm due to its excellent exploitation and exploration behavior in the identification of optimal weight score from the solution space, without allowing the solutions to stick in the local optimum. Finally, the fused features are fed into the Layered k-SVM (k-support vector machine) classifier for the recognition of fake face. The experimental results of our proposed approach are evaluated on three traditional benchmark face spoofing databases, namely the Replay-Attack, the CASIA Face Anti-Spoofing, and the MSU Mobile Face Spoof database. The outcome of our proposed approach exhibited steady and robust performance across all the three datasets. More commonly, our proposed approach executes well in the inter database tests and yields high performance, even though when only operated with minimized training data.