CAE-Net: Generalized Deepfake Image Detection using Convolution and Attention Mechanisms with Spatial and Frequency Domain Features
Effective deepfake detection is becoming crucial to address the growing usage of deepfakes in unethical practices. However, developing an accurate universal detection method is challenging due to the wide range of deepfake generation techniques. This study utilizes a diverse deepfake image dataset, provided by the 2025 IEEE Signal Processing Cup organizing committee as part of the DFWild-Cup competition, which exhibits significant class imbalance. To this end, we proposed a disjoint set-based multistage training method, utilizing an architecture, CAE-Net, that consists of a convolution- and attention-based weighted ensemble network. It employs three different architectures: EfficientNet, Data-Efficient Image Transformer, and ConvNeXt with wavelet transform. The model achieves an excellent 94.63% accuracy on the competition’s validation dataset, with a 97.37% AUC score. We visualize the focus regions of these models using Grad-CAM and demonstrate the effectiveness in grouping real and fake images into cohesive clusters using t-SNE plots. The model’s robustness against adversarial perturbations is also tested, showing inherent defensive properties of the ensemble approach.
Check out the work here: