DeepFakeDetector: A Multi-Branch Fusion Framework for Cross-Generator Detection of AI-Generated Images

Abstract

AI-generated imagery from diffusion and GAN-based models is now photorealistic, enabling misinformation, fraud, and non-consensual synthetic media. A key challenge for detectors is cross-generator generalization: models that overfit to generators seen during training often fail on new architectures. We present DeepFakeDetector, a multi-branch framework that combines complementary forensic cues: (i) a fine-tuned Vision Transformer (ViT) for global semantic consistency, (ii) a distilled DeiT model for efficient inference, (iii) a transfer-based EfficientNet-B0 CNN baseline for robust convolutional features, and (iv) a Gradient Field CNN that exploits structure-tensor coherence in image gradients. On an OpenFake subset, individual branches range from 72.69% (frozen ViT) to 87.90% (EfficientNet-B0) accuracy. We further evaluate on the AI-GenBench benchmark [1] and report fusion baselines via logistic regression stacking and a meta-classifier. Code: https://github.com/McMasterAISociety/DeepFakeDetector Models: https://huggingface.co/DeepFakeDetector Demo: https://www.deepfake-detector.app/.