Gender Obfuscation through Face Morphing

Shunxin Wang, Una M. Kelly, Raymond N. J. Veldhuis

Abstract: With the advancement of machine learning, facial biometric data has been widely adopted for person recognition. Soft biometrics such as gender, age and ethnicity can be extracted automatically from the facial photographs without permission. This raises privacy concerns since such auxiliary information might be utilized improperly. In this work, we apply face morphing to obfuscate gender information in face images, so that gender classifiers can no longer accurately predict gender, but the resulting images can still be used for identity verification. We further explore the reversibility of our approach and the results show that gender obfuscated through face morphing cannot be recovered or retrieved easily. Our approach is especially useful for an identity verification system which is sensitive to morphing attacks.