MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202641109034 A) filed by Easwari Engineering College; and Srm Institute Of Science And Technology, Ramapuram Campus on September 11, 2026, for A Transformer-Based Framework For Prompt Injection Attack Detection Using Fine-Tuned Distilroberta.
Inventors include Harini K; Vaishnavi Iyer S; and Dr. S Vanaja.
The application for the patent was published on September 18, 2026, under issue no. 38/2026.
Abstract: The present invention provides a system and method for AI-based prompt injection detection and prevention in Large Language Model (LLM) applications. The system receives user-submitted prompts and processes them through a normalization module to standardize textual representations and minimize variations that may conceal malicious instructions. An obfuscation restoration module identifies intentionally modified prompt content using a trained sequence-to-sequence language model and restores it into a semantically meaningful form. The restored prompt is then analyzed by a trained prompt injection detection model to identify adversarial instructions, attempts to override system instructions, bypass safety restrictions, manipulate LLM behavior, or obtain unauthorized information. A safety decision module classifies the prompt as safe or unsafe based on the detection result. Safe prompts are forwarded to the LLM for response generation, while unsafe prompts are blocked before reaching the LLM. The system displays either the generated response or an appropriate blocking notification, providing layered protection against direct and obfuscated prompt injection attacks. Total Number of Words in Abstract: 150
Disclaimer: Curated by HT Syndication.
