Kernekoncepter
A robust model can effectively detect AI-generated faces from a variety of GAN and diffusion-based synthesis engines, even at low resolutions and compressed image qualities.
Resumé
The content describes the development and evaluation of a model to distinguish real from AI-generated faces. The key points are:
- The model is trained on a large dataset of 120,000 real LinkedIn profile photos and 105,900 AI-generated faces from various GAN and diffusion synthesis engines.
- The model, based on the EfficientNet-B1 architecture, achieves a true positive rate (TPR) of 98% in classifying AI-generated faces, with a fixed false positive rate (FPR) of 0.5%.
- The model generalizes well to AI-generated faces from synthesis engines not seen during training, achieving an 84.5% TPR.
- The model is robust to low image resolutions, down to 128x128 pixels, and JPEG compression, maintaining high TPR even at low quality levels.
- Analysis suggests the model has learned a structural or semantic-level artifact in AI-generated faces, rather than just low-level artifacts.
- The model outperforms previous work focused on detecting GAN-generated faces and is more resilient to laundering attacks.
- The model provides a powerful tool in the adversarial battle against the misuse of AI-generated content, especially for detecting fake online profiles.
Statistik
The model achieves a true positive rate (TPR) of 98% in classifying AI-generated faces, with a fixed false positive rate (FPR) of 0.5%.
For faces generated by synthesis engines not used in training, the TPR drops to 84.5% at the same FPR.
The TPR remains high, 91.9%, even when classifying 128x128 pixel images.
The TPR degrades slowly, from 94.3% at JPEG quality 80 to 88.0% at quality 60.