”DOES YOUR MOBILE SUIT YOUR SKIN?”: ADDRESSING SKIN TONE DISPARITIES IN PRESENTATION ATTACK DETECTION FOR ENHANCED INCLUSIVITY OF SMARTPHONE SECURITY

Anudeep Vurity, Emanuela Marasco

Keywords: Skin Tone, Fairness, Mobile Biometrics, Finger Photo, Presentation Attack Detection, Presentation Attack, Color Spaces

TL;DR: Our paper examines how skin tone disparities impact Presentation Attack Detection (PAD) systems by introducing a novel ColorCubeNet framework by processing multi-color channel data to improve inclusivity and fairness in mobile biometric security.

Abstract:

Mobile devices are at a heightened risk for cybercrime due to the sensitive personal and financial data they handle. Biometric authentication provides a robust,convenient, and secure way to protect smartphones by using unique user characteristics like fingerprints, facial features, or voice patterns for access. Existing mobile biometric technology often relies on RGB cameras to capture biometric
samples, such as face images or finger photos, making them vulnerable to spoofing (e.g., 3D masks, display or printout attack). The security of these systems is effectively addressed by integrating a Presentation Attack Detection (PAD) module. Existing PAD solutions do not account for diverse physical characteristics such as skin tone. As a result, marginalized groups face higher misidentification
rates or false rejections, reducing access to services and increasing security risks.

This paper introduces a novel deep learning framework called ColorCubeNet that is designed to process ColorCube, a multi dimensional data representation by combining information from RGB, HSV and YCbCr color spaces. This data cube leverages the joint capabilities of RGB, HSV, and YCbCr color spaces to depict color more sophisticatedly. By incorporating features from multiple complementary color channels, this approach can effectively handle a variety of skin tones. We utilized three EfficientNet-B0 models, each trained on ImageNet using RGB, HSV, and YCbCr color spaces, and then fine-tune them on the ColorCube representation to fully exploit the combined information from all three color spaces. Additionally, a channel-attention mechanism is integrated into the architecture, enabling the extraction of key features from different input channels and exploit their combined performance. Results show that the proposed approach outperforms traditional RGB methods by reducing skin tone disparities by 50%.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11869