anonymous-author-sub/seeable
📦 Open Source Projectanonymous-author-sub
Advanced deepfake detection framework using soft discrepancies and bounded contrastive learning for superior forgery exposure.
SeeABLE represents a significant step forward in digital forensics for AI-generated media. Unlike traditional deepfake detectors that rely on rigid classification boundaries, SeeABLE introduces a methodology centered on 'soft discrepancies'—subtle, localized artifacts and inconsistencies that characterize synthetic forgeries. The framework leverages bounded contrastive learning to map features into a latent space where real and fake samples are more effectively separated, even when the forgery is highly realistic. This approach mitigates the risk of overfitting to specific training datasets and improves generalization across different generative architectures. The repository provides the necessary Python-based infrastructure for researchers to implement these detection techniques, offering a modular pipeline for feature extraction and contrastive training. It is particularly effective at identifying artifacts that standard CNN-based detectors might overlook, making it a valuable asset for security researchers and media integrity verification teams.
💡Highlights
- ├─Bounded contrastive learning
- ├─Soft discrepancy analysis
- └─Robust forgery detection
🎯For
- ├─AI Security Researchers
- └─Digital Forensics Experts