Problem
Fake accounts increasingly use faces from image generators as profile photos, and members cannot tell them from real photos by eye. A detector for a network with hundreds of millions of daily users must catch most synthetic photos and rarely flag a real member. Generic forensics models lose accuracy on the newest face generators.
Example from the article: LinkedIn's Trust Data team and a Berkeley forensics lab built a detector for StyleGAN faces and tested it on 100,000 real LinkedIn profile photos.
Solution in brief
GAN faces are cropped and aligned the same way, so a low-dimensional model trained only on synthetic faces reconstructs them well and reconstructs real photos badly. The 128-dimensional embedding of that model feeds a logistic regression, which flags 99.6% of StyleGAN faces at a 1% false positive rate.
Questions this case answers
- How to detect fake images produced by a known family of generators with a lightweight model when real images are diverse and synthetic ones are few
- How to choose a detection threshold under a fixed false positive budget when a wrong flag hurts legitimate users
- How to detect AI-generated profile photos used for fake accounts?
- How to build a cheap face-forensics detector when you have only a few thousand synthetic faces?
Tools and methods
- PCA (128 components) — learns a linear face basis from synthetic faces only; reconstruction error or the 128 coefficients are the features.
- Autoencoder (3 layers, 128-unit hidden layer) — a nonlinear version of the same idea, with an average of eight points higher recall than PCA on reconstruction error alone.
- Logistic regression — classifies the 128-dimensional embedding and gives the headline result.
- Fourier low-pass embedding — a fixed baseline that shows a generic embedding is not enough.
- StyleGAN 1/2/3, Stable Diffusion — generators used to build the synthetic training and test sets.
Google or a one-time email code. No password.