An Efficient Pre-Treatment and Machine Learned Framework for Detection of Spoofed Faces

• Published in: 2024 2nd International Conference on Emerging Trends in Engineering and Medical Sciences (ICETEMS) pp. 878 - 884
• Main Authors: Chakole, Vijay V., Dixit, Swati R., Karule, P. T., Agarkar, Poonam T., Palsodkar, Prachi, Palsodkar, Prasanna
• Format: Conference Proceeding
• Language: English
• Published: IEEE 22.11.2024
• Subjects: authentic and synthetic faces; Biological system modeling; conventional descriptors; Faces; Feature extraction; Filtering; Machine learning; modified difference of Gaussian filtering; Personal identification systems; Real-time systems; replay attack; Security; Supervised learning; Support vector machines; SVM
• DOI: 10.1109/ICETEMS64039.2024.10965062

Personal identification systems have been an intense need to provide security in several real-time applications that primarily include workplace access, online transactions, criminal forensics, law enforcement, etc. for either identifying the authentic individual or subject identification. The face anti-spoofing system (FAS) presented in this article uses a novel pre-treatment with a machine-learning (support vector machine (SVM)) approach to distinguish authentic and synthetic faces (Replay attack (RA)) generated from publicly available IDIAP data stores. A dual-branch image pre-treatment stage not only improves the quality of the image but also enhances the edge details using the modified difference of Gaussian filtering (DoG), antialiasing filter, and illumination correction. The resultant mean image from both operations produces a distilled image that is used to extract the concerned region and extricate representative features. The rich facial attributes are represented using diverse conventional descriptors that not only include the textural aspects but also the facial liveliness. Thus a single face is represented using seven descriptors amounting to more than 2000 real values. Experiments conducted on 13950 (4000:9950) samples from both categories with a 75:25% train: test ratio revealed that the classifier outperformed other competing approaches with the highest classification accuracy of 99.534%. The proposed face antispoofing framework possesses generalization ability when tested over real samples.