Exploring adversarial deep learning for fusion in multi-color channel skin detection applications

Abstract

Deep learning, a robust framework for complex learning, outperforms previous machine learning approaches and finds widespread use. However, security vulnerabilities, especially in fusion in multi-color channel skin detection applications using adversarial machine learning (AML) and generative adversarial networks (GANs), lead to misclassifications. Researchers are actively exploring AML's and GANs' impact on misclassification, focusing on vulnerabilities in lighting conditions, skin-like patches in lesion segmentation, and insufficient data in facial emotion recognition. Yet, these areas only scratch the surface of potential AML vulnerabilities and GANs. To comprehensively address challenges, an in-depth investigation into AML and GANs components is crucial to uncover underlying reasons for misclassifying skin detection. This study addresses challenges of fusion in multi-color channel skin detection by creating a diverse dataset with 17M patches for enhanced feature fusion/training and meeting dataset criteria, investigating misclassifications using various deep learning models belonging to AML and GANs and color spaces (e.g., RGB, YCbCr, HSV, YUV), and exploring binary and multiclass scenarios. Notably, YCbCr outperformed RGB, achieving 98 % for binary skin classification, 84 % and 69 % for multiclass four and five-class scenarios. Binary classification for skin tones and their skin-like counterparts (e.g., black skin tone and black skin-like) yielded 97 %, 81 %, 60 %, and 51 % for black, brown, medium, and fair, respectively. Exploration of darker skin tones showed improved accuracy. Benchmarking with a CNN and RNN hybrid achieved 99 % accuracy, surpassing the initial 91 %, while SAE reached 97 %. The study explores implications of overlapping between skin and skin-tone recognition, offering insights for developing a generalized skin detector. The investigation demonstrates that improper color space selection can make lighting conditions exploitable in AML attacks and GANs, emphasizing the crucial role of color space choice in mitigating vulnerabilities.