Dual-Branch Ensemble for Synthetic Image Attribution

A research article introduces a hybrid model for Synthetic Image Source Attribution (SIA) designed to tackle distribution changes from ideal training images to real-world images that have undergone unknown post-processing, such as JPEG compression and blurring. This framework, presented for the DLMMDD Challenge at ICANN 2026, combines a semantic component utilizing EfficientNet-B0 with Exponential Moving Averaging and Label Smoothing, alongside a forensic component that extracts 126 mathematical features (including SVD spectral profiles and Local Binary Patterns) from high-pass noise residuals. These features are then compressed using Truncated SVD and classified with XGBoost. The method was tested on a dataset comprising 10 generators, with 55% of the test set being degraded, aiming for reliable attribution.