Deepfake Image Detection Using Spatial Rich Model Features and Deep Learning Fusion

Abstract

We present a spatial rich model (SRM)-based fusion approach for accurate and reliable deepfake image detection in a binary real-versus-fake facial image classification setting. The proposed method combines the effectiveness of modern deep learning models with residual-oriented forensic information extracted through SRM-related processing. First, we evaluate multiple baseline architectures, including EfficientNetV2-S, ResNeXt50, ViT-B/16, and a lightweight custom CNN, under the same validation protocol. Then, we select EfficientNetV2-S as the main RGB backbone and extend it with an additional SRM-related branch, fused into the network at different depths. We show that the proposed SRM-based fusion models outperform the strongest RGB-only baseline. Finally, we further improve the results by combining the strongest fusion variants with selected baseline models using soft voting. The proposed ensemble achieves 97.48% accuracy, 96.46% precision, 98.57% recall, and 97.50% F1-score on the global test set, showing that SRMrelated forensic cues can effectively complement deep visual representations for deepfake image detection.