SCADET: A detection framework for AI-generated artwork integrating dynamic frequency attention and contrastive...

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Generative AI advancements create challenges in verifying AI-generated image authenticity.
• Distinguishing AI-generated images from real ones is crucial for digital media integrity.
• Existing detection methods struggle with diverse AI generation techniques and artistic styles.

Purpose of the Study:

• To propose SCADET, a novel detection framework for AI-generated images.
• To enhance authenticity verification and originality validation of digital content.
• To improve the generalization capabilities of AI-generated image detection across different models.

Main Methods:

• SCADET integrates Dynamic Frequency Attention Network (DFAN) for adaptive frequency domain analysis.
• DFAN dynamically adjusts attention based on image artistic styles.
• Contrastive Spectral Analysis Network (CSAN) uses contrastive learning to build discriminative feature spaces.

Main Results:

• SCADET achieved AUC values of 0.962 (full image) and 0.801 (local image) on the AI-ArtBench dataset.
• Demonstrated substantial improvements of 30.5% and 34.4% over baseline methods.
• Maintained stable cross-model performance with average accuracy of 0.81 and low variance across various generation techniques.

Conclusions:

• SCADET effectively detects AI-generated images, addressing authenticity and originality challenges.
• The framework's components, DFAN and CSAN, are validated for their effectiveness.
• Results advance AI-generated content detection and offer insights for digital media authenticity.