Seeing Through Deepfakes: Explainable AI (XAI) for Image Forensics and Visual Detection

Deepfake technology has made it increasingly easy to fabricate highly realistic facial videos. This creates serious challenges for individual privacy and information security. Many existing deep fake detection models based on Deep Convolutional Neural Networks (DCNNs) perform well on benchmark datasets, but their performance is not very good on real-world content, particularly for faces from underrepresented demographics. In practice, these models often misclassify Indian, South Asian, and other Asian faces with high false positive rates. To address this short coming, we design a deepfake detection framework tailored to Indian users, built on a strong pre-trained DCNN that is adapted to the target domain using parameter-efficient fine-tuning. We have created a new proprietary dataset, DF-India, consisting of diverse Indian facial identities and realistic capture conditions, including variations in device quality, resolution, compression, and common online distortions, to better reflect deployment scenarios. The framework also incorporates an Explainable AI (XAI) component based on Gradient-weighted Class Activation Mapping (Grad- CAM), which highlights the image regions driving the model's decisions. Experimental results indicate that the proposed approach offers improved detection accuracy, stronger cross- domain robustness, and reduced demographic disparity in error rates, leading to more trustworthy deepfake detection for South Asian populations.