🤖 AI Summary
This work addresses the challenge of distribution shift and degraded source-tracing accuracy in synthetic images caused by unknown post-processing operations such as JPEG compression or blurring. To this end, the authors propose a dual-branch ensemble framework that efficiently integrates semantic deep learning—using EfficientNet-B0 enhanced with exponential moving average (EMA) and label smoothing—with lightweight mathematical forensic features derived from high-pass noise residuals. These handcrafted features comprise a 126-dimensional vector incorporating singular value decomposition (SVD) spectra, local binary patterns, and dimensionality reduction via truncated SVD. The combined representations are fused through an XGBoost classifier. Requiring only CPU computation, the method achieves 95.60% accuracy on a challenging benchmark where 55% of test samples undergo degradation, completing the entire pipeline within 6.5 hours. This approach demonstrates remarkable robustness to post-processing while offering high accuracy, computational efficiency, and practical deployability.
📝 Abstract
The rapid advancement of text-to-image (T2I) models has necessitated robust Synthetic Image Source Attribution (SIA) methodologies. A critical challenge in SIA is the distribution shift between pristine training images and real-world deployed images, which undergo unknown post-processing operations such as JPEG compression and blurring. In this work, proposed for the DLMMDD Challenge at ICANN 2026, we introduce a dual-branch ensemble framework fusing Semantic Deep Learning with Mathematical Forensic Feature Extraction. The semantic branch employs EfficientNet-B0 regularized with Exponential Moving Averaging (EMA) and Label Smoothing. The forensic branch extracts 126 mathematical features -- including SVD spectral profiles and Local Binary Patterns -- from high-pass noise residuals, compressed via Truncated SVD and classified with XGBoost. Evaluated on a dataset of 10 generators where 55% of the test set is degraded, our approach achieves a private leaderboard accuracy of 95.60%. Furthermore, the entire pipeline is highly computationally efficient, requiring no GPU acceleration and executing end-to-end on a standard CPU in under 6.5 hours, highlighting the practicality and scalability of mathematical forensics for real-world deployment.