Face Detection and Recognition Under Real-World Scenarios - Dealing with Deepfake Incidents and Malicious Data Distortions

The growing use of deepfake technology and synthetic facial images presents significant challenges to biometric verification systems, particularly when identity-specific features are obscured by textured masks, halftones, and watermarks. This paper evaluates six face detection methods and four state-of-the-art facial recognition engines on two types of data: unaltered facial images and synthetic faces masked with diverse textures. We proposed a custom Python library for generating image distortions and obfuscations to simulate textural noise and assessed their impact on face verification. Results revealed vulnerabilities of current systems to textured masks and watermarks, highlighting challenges posed by occluded features in manipulated media. Additionally, we introduced a novel face detection method utilizing this augmentation library, achieving up to 99% detection accuracy under noisy conditions. Our analysis identified statistically significant obfuscation techniques that affect model performance the most, providing insights for improving robustness against real-world distortions.