MUMBAI, India, July 30 -- Intellectual Property India has published a patent application (202641088210 A) filed by Anjani Suputri Devi D; Geesala Veerapandu; Suneetha Eluri; Chinnam Sabitha; D. Sasi Rekha; M. Kishore Kumar; and D. Durga Bhavani on July 20, 2026, for Breaking The Illusion: A Hybrid Deep Learning Approach For Detecting Ai And Human Generated Text.
Inventors include Anjani Suputri Devi D; Geesala Veerapandu; Suneetha Eluri; Chinnam Sabitha; D. Sasi Rekha; M. Kishore Kumar; and D. Durga Bhavani.
The application for the patent was published on July 24, 2026, under issue no. 30/2026.
Abstract: In recent times Deep Learning has transformed various fields such as computer vision, speech recognition, and Natural Language Processing (NLP). One of its crucial applications is fake content detection, including deepfake videos, manipulated images, and AI-generated text. With the advancement of large language models like GPT, AI-generated text has become highly sophisticated, making it difficult to differentiate from human-written content. This poses significant challenges in journalism, social media, and online communication, where AI-generated misinformation can impact public opinion and trust. Among these challenges, deepfake text detection is critical, as AI-generated text can be used for fake news, misinformation, and online deception. The primary challenge is to develop a model that accurately classifies real and fake text while reducing false positives in human- written content.Existing systems attempt to address fake news detection using techniques like Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTMs), Bidirectional Long Short-Term Memory (BiLSTMs), and Transformer models. RNNs suffer from vanishing gradient problems, making them less effective for long texts, and they struggle with capturing long- range dependencies. LSTMs improve upon RNNs by maintaining longer memory, but they still face limitations when dealing with large- scale datasets due to high training time and computational complexity. BiLSTMs enhance LSTMs by processing text in both forward and backward directions, improving contextual understanding. However, they require significantly more computational power, struggle with long sequences due to memory constraints, and can still suffer from vanishing gradient issues in extremely long texts.The proposed system integrates BERT, RoBERTa, and DeBERTa to enhance deepfake text detection. BERT and RoBERTa provide strong contextual understanding and have been widely used for text classification tasks, while DeBERTa further improves upon these models by incorporating disentangled detection, which seperates word meaning and position for better contextual understanding and classification performance. DeBERTa will work well with superior generalization and efficiency. DeBERTa-based fake text detection starts by collecting real and AI-generated text from various sources. The text is then preprocessed by removing unnecessary words, tokenizing, and converting it into a numerical format for analysis. The model is trained to recognize patterns using its advanced attention mechanism, which helps understand context and detect fake content. Finally, it classifies text as real or fake, making it useful for fact-checking tools, social media monitoring, and news verification.
Disclaimer: Curated by HT Syndication.
