Semi Adversarial Networks in Matlab

Abstract:

Recent research has established the possibility ofdeducing soft-biometric attributes such as age, gender, and racefrom an individual’s face image with high accuracy. However,this raises privacy concerns, especially when face images collectedfor biometric recognition purposes are used for attribute analysiswithout the person’s consent. To address this problem, we developa technique for imparting soft biometric privacy to face imagesvia an image perturbation methodology. The image perturbationis undertaken using a GAN-based Semi-Adversarial Network(SAN) — referred to as PrivacyNet — that modifies an inputface image such that it can be used by a face matcher formatching purposes but cannot be reliably used by an attributeclassifier. Further, PrivacyNet allows a person to choose specificattributes that have to be obfuscated in the input face images(e.g., age and race), while allowing for other types of attributes tobe extracted (e.g., gender). Extensive experiments using multipleface matchers, multiple age/gender/race classifiers, and multipleface datasets demonstrate the generalizability of the proposedmulti-attribute privacy enhancing method across multiple faceand attribute classifiers.