Key points are not available for this paper at this time.
The rise of deepfake technology has transformed media synthesis, enabling the creation of hyperrealistic yet manipulated images and videos. Although these innovations offer creative opportunities, they also introduce severe risks such as misinformation, identity theft, and decreased trust in digital content. This study presents a hybrid approach to deepfake image detection that integrates features from the spatial and frequency domains to improve detection accuracy. The proposed method combines multiscale Convolutional Neural Networks (CNNs), frequency domain analysis, attention-based transformer networks, and ensemble learning to identify manipulation artifacts and enhance classification robustness. The model was tested on a large-scale dataset of 140,000 images, evenly divided between real and fake images, achieving a training accuracy of 93% and a testing accuracy of 88%. Using adversarial training and advanced feature extraction techniques, the proposed approach effectively detects subtle artifacts introduced during manipulation. The experimental results illustrate the model's ability to generalize across diverse manipulation methods and datasets, making it a scalable and reliable tool for real-world applications. This research emphasizes the significance of hybrid detection frameworks in addressing the complexities of synthetic media forensics and underscores the necessity of multidomain feature integration to address the growing challenges posed by deepfakes.
Yadav et al. (Wed,) studied this question.