Boosting Geometric Invariants for Discriminative Forensics of Large-Scale Generated Visual Content

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Forensics

Background:

• Generative AI creates realistic images, challenging traditional forensic methods.
• Current forensic models lack robustness and interpretability for AI-generated content.
• Empirical learning algorithms struggle with high demands for trustworthy forensic tasks.

Purpose of the Study:

• To develop a robust and interpretable method for forensic analysis of large-scale AI-generated images.
• To enhance the discriminability of classical geometric invariants for digital forensics.
• To address limitations of empirical learning in AI image forensics.

Main Methods:

• Extended classical geometric invariants to a hierarchical convolutional neural network architecture.
• Utilized overcompleteness for automatic selection of task-discriminative features.
• Applied boosted invariants to diverse generative models including GANs and diffusion models.

Main Results:

• Achieved state-of-the-art discriminability against large-scale content diversity in generated images.
• Demonstrated high efficiency on training examples and intrinsic invariance to geometric variations.
• Showcased improved interpretability in the forensic process compared to traditional methods.

Conclusions:

• Boosted geometric invariants offer a robust, interpretable, and highly discriminative solution for AI-generated image forensics.
• The proposed method effectively handles diverse generative models and large datasets.
• This approach advances trustworthy AI by improving the reliability of digital forensic analysis.