Primary Investigator:
Hany Abdel-Khalik

Project Members
1. Raghav Aggarwal 2. Dr. Arvind Sundaram 3. Dr. Hany S Abdel-Khalik

This research introduces a framework for Cognizant Deep Fakes (CDF), where Large Language Models (LLMs) are used to generate synthetic text carriers for secure information transfer. By integrating the Llama-zip compression algorithm into the token-selection process of a Mistral 7B model, we create "deep fake" text that is statistically indistinguishable from standard AI outputs yet remains "cognizant" of an embedded data payload. Unlike traditional methods, CDF leverages the high-dimensional probability space of LLMs to ensure robustness against detection. This study validates an end-to-end CDF pipeline, demonstrating a 100% recovery rate of embedded data through a custom web-based interface, providing a novel approach to secure communication in denied environments.