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

Authors: Orazio Pontorno, Mattia Litrico, Luca Guarnera, Mario Valerio Giuffrida, Sebastiano Battiato

Published: 2026-06-29 16:31:57+00:00

Comment: Accepted at the European Conference on Computer Vision (ECCV) 2026

AI Summary

μFlow is a one-class deepfake detector for faces trained exclusively on real images, without relying on pseudo-deepfakes or synthetic artifacts. It leverages the observation that averaging multiple images amplifies consistent generative traces, creating highly discriminative feature representations. The method models the distribution of features from averaged images using a normalizing flow to align the feature space of individual images, leading to a likelihood-based criterion for distinguishing real from fake samples with strong generalization.

Abstract

Current generative models, including GANs and diffusion models, have reached an outstanding level of photorealism, posing significant risks to privacy and security. To ensure real-world applicability, deepfake detectors must generalise effectively to unseen generators. However, most existing approaches rely on supervised training with both real and fake images, which limits their generalisation especially across generators categories (e.g. GANs vs DMs). In this work, we introduce $μ$Flow, a one-class deepfake detector trained only on real images without relying on pseudo-deepfakes or synthetic artifacts. Our approach builds on the observation that averaging multiple images amplifies consistent generative traces, producing highly discriminative feature representations. We leverage this property by modelling the distribution of features extracted from averaged images and training a normalizing flow to align the feature space of individual images with this distribution. This alignment yields a likelihood-based criterion that separates real and fake samples while promoting strong generalisation. We evaluate $μ$Flow on a fully out-of-distribution setting, where both real and fake datasets are unseen during training. Experimental results show that our method significantly outperforms SOTA detectors. Project page: https://opontorno.github.io/MuFlow.

Key findings

μFlow significantly outperforms state-of-the-art deepfake detectors in a fully out-of-distribution setting, especially when generalizing across different generator categories (GANs vs. DMs). It achieves an average accuracy of 94.7%, surpassing the second-best performing approach by +2.1% even when compared to methods trained on mixed datasets of GANs and DMs, and shows robust performance against various content-preserving transformations and signal-degrading perturbations.

Approach

μFlow operates as a one-class classifier, trained solely on real images. It builds a discriminative latent space by extracting and averaging features from multiple real images, then fitting a Gaussian Mixture Model to this representation. A normalizing flow (FastFlow) is subsequently trained to map features from individual real images into this learned discriminative space, with test images classified as fake if their likelihood in this space is low.

Datasets

FFHQ, CelebA-HQ, WILD (including GANs, DM-OS, DM-CS)

Model(s)

ResNet50 (as frozen encoder), FastFlow (normalizing flow)

Author countries

Italy, UK

← Previous