Hybrid Semantic and Spectral Ensemble for Robust Synthetic Image Source Attribution

This paper presents a dual-branch ensemble for synthetic image source attribution under distribution shifts caused by real-world post-processing. Its semantic branch uses EfficientNet-B0 with exponential moving averaging and label smoothing, while its forensic branch extracts 126 features from high-pass noise residuals, including SVD spectral profiles and Local Binary Patterns, then applies truncated SVD and XGBoost. On a dataset covering 10 generators, with 55% of test images degraded, the method achieved 95.60% accuracy on a private leaderboard. The complete pipeline reportedly runs without GPU acceleration and finishes on a standard CPU in under 6.5 hours.