Generative Modeling towards Model-Free Kinship Verification Attack

Abstract Visual kinship understanding has raised privacy concerns in the past few years. That being said, the kinship relationship can be simply detected and exposed to the public, with or without the permission of photo owners, leading to severe privacy leaks or malicious retaliation against families. Visual kinship verification, a procedure that determines whether two people have kin relations through facial images, is the primary concern among kinship applications. To that end, this paper explores practical approaches to compromising CNN and Transformer based deep kinship verification models, to protect users' kinship privacy. Specifically, two generative models, Kinship-advGAN and Kinship-advTransGAN, are proposed to generate small perturbation vectors in face images to flip the verification results. To validate our methods, we conduct extensive experiments on visual kinship benchmarks, including FIW, KinFaceW datasets, and Kaggle Kinship Verification Challenges. The results show that our methods outperform several state-of-the-art methods and reveal vital insights: the CNN-based method attains a higher success rate of attacks, while the Transformer-based method provides more decentralized adversarial attacks on kinship verification.