DeFaX: A Cross-Attention Fusion Framework for Robust and Explainable Deepfake Detection

The increasing realism of GAN-generated facial forgeries has intensified the need for reliable deepfake detection systems capable of operating under diverse generative conditions. Traditional detectors often struggle to simultaneously capture global semantic structures and fine-grained local artifacts, thereby limiting their robustness when confronted with previously unknown sources. To address this challenge, this study introduces DeFaX, a cross-attention fusion framework that integrates the hierarchical global reasoning of the Swin Transformer with the localized feature extraction capability of EfficientNet through the SwinEffAttn module. The framework is further enhanced by the inclusion of Grad-CAM and LIME, which facilitate post-hoc interpretability and provide a visual justification for model decisions. DeFaX also incorporates a lightweight classification head and is demonstrated through a web-based interface to…