APTIMA INC — Department of Defense SBIR Phase I: N202-126

Description

Information and influence operations (IO) have rapidly become a de facto tool in near-peer conflict, as evidenced by the Russian invasion of Crimea, the 2016 US Presidential Election, and the COVID-19 response. Presently, the Department of Defense (DoD) relies on manual labor to generate text that simulates IO tactics, techniques, and procedures, which does not scale to enable real-time text injection during training exercises. The DoD needs a scalable capability that can simulate IO TTPs by generating realistic text and injecting it into a training exercise in real time. In response to this need, Aptima, Inc., and our partner, Dr. Janelle Ward, propose to develop CHARIOT: Charlie Harnessing Attention for Realistic injects in Information Operations Training. CHARIOT leverages advances in transformer-based generative language models to simulate IO TTPs by providing real-time injects to military training exercises. CHARIOT advances the state-of-the-art in IO simulation by (1) using intelligent fine-tuning and prompting techniques on Aptima’s Charlie infrastructure to create multiple distinct personas that inject realistic and varied texts into the training scenario, (2) providing an intuitive and naturalistic user experience that allows as few as two or three White Cell operators to control the function of these generative language models, and (3) integrating generative capabilities with Aptima’s YAADA™ data science platform to enable performance analysis and support more flexible model training to fine-tune personas and behaviors. When fully realized, the CHARIOT capability will produce text injects that vary in size (from approximately 50 to 800 words), represent diverse perspectives in the operational environment, and are generated at a volume and velocity comparable to real-world IO TTPs (e.g., 50,000 50-word texts/hour, 1,000 800-word texts/hour). These advances will improve the rigor and robustness of US military training and prepare warfighters for effective and robust response to adaptive IO threats.