VISION TRANSFORMER- BASED FRAMEWORK FOR ROBUST DETECTION AND CLASSIFICATION OF AIGENERATED IMAGES AND VIDEOS

The darker side of the impressive progress of generative artificial intelligence is the fact that the authenticity of digital media, cybersecurity, and social trust are posed with the most serious challenges due to the improved realism of synthetic images and videos. Conventional convolutional neural network founded detection models are generally sensitive to local spatial artifacts, which are becoming less effective with high fidelity generative models like GANs and diffusion based networks. To overcome these shortcomings, this paper suggests a framework of Vision Transformer to allow robust detection and classification of images and videos generated by AI. The proposed framework incorporates the global self attention mechanisms to learn long range spatial dependencies, semantic inconsistencies, and cross region correlations that can hardly be learned with the conventional CNNs. Moreover, a hybrid spatio temporal module is introduced to improve the video forgery detection by performing the process of modelling frame level spatial artifacts and temporal coherence anomalies. Frequency domain feature embedding and artifact preserving augmentation technique are included to enhance the cross dataset generalization and stability to post processing algorithms like compression and resizing. Moreover, the regularization of adversarial consistency is also presented to enhance resistance to adversarial perturbation. Experimental assessments conducted on benchmark sets proves that the proposed framework has high detection accuracy and better generalization than the state of the art methods. The findings indicate the usefulness of Transformer driven architectures to overcome the emerging challenges in the field of AI generated content forensics and offer a long-term scalable base to provide secure deployment on a real-world basis..