Effect of generator-label supervision on binary deepfake detection

Determine whether incorporating generator identity labels into the training of deepfake image detectors—by formulating training as multiclass classification with one real class and one class per generator—consistently improves binary real-versus-synthetic detection performance compared to direct binary training.

Background

Deepfake detection is commonly treated as a binary classification task (real vs. synthetic), even when datasets include labels for the specific generators that produced synthetic images. The paper investigates whether reframing training as a multiclass problem using generator identities can yield better binary detection performance.

To explore this, the authors evaluate plain multiclass training with different strategies to fuse multiclass outputs into binary decisions (sum fusion and max fusion), as well as dual-head architectures that jointly optimize binary and multiclass losses. Despite empirical insights, the question of whether generator-label supervision inherently improves binary detection remains explicitly posed as an open question.