MUMBAI, India, Aug. 17 -- Intellectual Property India has published a patent application (202641096138 A) filed by Velammal Institute Of Technology; Mr. Raja R; Mr. Anbumani A; Dr. R Venkadesh; Nandhitha B; Priya Mahalakshmi S; Sandhya B; and Tejasri K on August 08, 2026, for Vision Transformer Based Framework For Detection And Classification Of Real And Ai-Generated Visual Media.
Inventors include Mr. Raja R; Mr. Anbumani A; Dr. R Venkadesh; Nandhitha B; Priya Mahalakshmi S; Sandhya B; and Tejasri K.
The application for the patent was published on August 14, 2026, under issue no. 33/2026.
Abstract: The present invention discloses a Vision Transformer (ViT) based framework for the automated detection and classification of visual media as either real (human- created) or AI-generated (synthetic/deepfake). The framework addresses critical limitations of prior Convolutional Neural Network (CNN) based detection approaches, specifically their inability to model global spatial dependencies and their poor generalisation across diverse deepfake generation techniques.The system processes input images by dividing them into non-overlapping fixed-size patches, which are linearly projected into embedding vectors with positional encodings. These embeddings are processed by a multi-layer transformer encoder employing multi-head self-attention mechanisms, which capture both local and global inter-patch relationships. A learned classification token accumulates global visual context and is passed through a classification layer to produce binary class probabilities. Test-time augmentation comprising horizontal flipping and centre-cropping is applied to enhance inference robustness through probability averaging.For video media, a two-stage forensic detection pipeline is implemented: direct temporal video analysis for motion physics and consistency verification, supplemented by frame-level image forensics on sampled key frames. The complete system is integrated within a modular pipeline encompassing data collection, preprocessing, patch embedding, feature extraction, classification, and output generation, and is deployed via a Flask-based web application supporting real-time detection with confidence scoring.The invention achieves superior detection accuracy, robustness, and cross-dataset generalisation compared to CNN-based prior art, making it applicable to digital forensics, media authentication, cybersecurity, law enforcement, and social media content moderation.
Disclaimer: Curated by HT Syndication.
