$μ$Flow: Leveraging Average Images for Improving Generalisation of Deepfake Faces Detectors

TL;DR

This paper proposes μFlow, a one-class deepfake detector trained only on real images that achieves strong generalization to unseen generative models.

Problem

State-of-the-art generative models (GANs, diffusion models) produce highly realistic images, making generalization a critical challenge for deepfake detectors. Existing supervised methods trained on both real and fake images suffer significant performance drops when facing new generator categories not seen during training.

Approach

Building on the observation that averaging multiple images amplifies consistent generative traces, the method models the distribution of features extracted from averaged real images. A normalizing flow is then trained to align the feature space of individual images with this distribution, yielding a likelihood-based criterion for separating real and fake samples.

Results & Contribution

Evaluated in a fully out-of-distribution setting (unseen real and fake datasets during training), μFlow significantly outperforms state-of-the-art detectors. The key contribution is achieving robust generalization to novel generators by training exclusively on real images.