Advanced deepfake image detection: A robust framework using Inceptionv3 and Xception

Deepfake technology, driven by generative artificial intelligence, has emerged as a significant challenge to digital trust by enabling the creation of hyper-realistic yet fraudulent media content. Such manipulations are increasingly used for misinformation, financial fraud, and cybersecurity attacks. To address this, the present study introduces a robust deepfake image detection framework based on Convolutional Neural Networks (CNNs), specifically InceptionV3 and Xception. A dataset of 19,457 images containing authentic and AI-generated forgeries was preprocessed through normalization, augmentation, and class balancing to enhance generalization. Both models were fine-tuned using transfer learning on stratified training, validation, and test sets. The results demonstrate strong detection capabilities, with InceptionV3 achieving 92.19% accuracy and Xception achieving 93.44%, while an ensemble approach further improved accuracy to 94.50%. Comparative evaluation against existing approaches indicates that the proposed framework outperforms several image-based detection systems in terms of accuracy, precision, and recall. This study shows that fine-tuned CNNs offer scalable and reliable solutions for detecting deepfakes, thereby supporting cybersecurity, media authentication, and digital forensics. Future work may extend this framework to multimodal and temporal deepfake detection for enhanced robustness against evolving threats.