Abstract: We propose a unified approach to simultaneously addressing the conventional setting of binary deepfake classification and a more challenging scenario of uncovering what facial components ...