ML Security PapersSelf-Supervised Learning for Detecting AI-Generated Faces as Anomalies
Mian Zou 1, Baosheng Yu 2, Yibing Zhan 3, Kede Ma 1
Published on arXiv
2501.02207
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
The EXIF-based anomaly detection framework generalizes to nine state-of-the-art generative models (GANs and diffusion models) without requiring synthetic training data from those generators.
AIGFD-EXIF
Novel technique introduced
The detection of AI-generated faces is commonly approached as a binary classification task. Nevertheless, the resulting detectors frequently struggle to adapt to novel AI face generators, which evolve rapidly. In this paper, we describe an anomaly detection method for AI-generated faces by leveraging self-supervised learning of camera-intrinsic and face-specific features purely from photographic face images. The success of our method lies in designing a pretext task that trains a feature extractor to rank four ordinal exchangeable image file format (EXIF) tags and classify artificially manipulated face images. Subsequently, we model the learned feature distribution of photographic face images using a Gaussian mixture model. Faces with low likelihoods are flagged as AI-generated. Both quantitative and qualitative experiments validate the effectiveness of our method. Our code is available at \url{https://github.com/MZMMSEC/AIGFD_EXIF.git}.
Key Contributions
- Self-supervised pretext task combining ordinal EXIF tag ranking (aperture, exposure time, focal length, ISO speed) and face manipulation classification to learn camera-intrinsic and face-specific features from real photos only
- Gaussian mixture model built on the learned feature distribution of photographic faces to flag AI-generated faces as low-likelihood anomalies
- One-class anomaly detection framing that generalizes to novel generators without retraining on synthetic images
🛡️ Threat Analysis
Proposes an AI-generated content detection method targeting synthetic face images produced by GANs and diffusion models — deepfake detection is explicitly listed under ML09 output integrity and content provenance.