DeepMedia: Rapid and Accurate DeepFake Detection to Counter Russian and Chinese Information Warfare

A DeepFake is an AI generated or manipulated video, audio, or news-article that appears 100% realistic to the human eye and can fool existing biometric detection methods. Russia and China will use this technology in a future peer-to-peer conflict with the US.  The Air Force currently lacks the manpower and critical language skills to identify DeepFake content at scale. The most effective way to identify a DeepFake is through audio, voice, and language queues. Native language speakers are the most effective way to identify DeepFakes. The Air Force suffers a shortage of skilled linguists in strategic threat languages - specifically Russian and Ukrainian. They can take 44 weeks for basic proficiency and 5-6 years for mastery. There are millions of sound bits and videos circulating across Publicly Available Information (PAI) sources and social media (Tik Tok, Instagram, Facebook). All of these contain synthetically manipulated DeepFake content and the number is growing exponentially. The Air Force simply does not have the numbers of native speakers or cryptologists to keep up with the flood of DeepFake content in PAI. This risks adversary DeepFake content proliferating across PAI and spreading disinformation which can reduce the speed and accuracy of Air Force tactical and operational decision-making. This ultimately weakens the Air Force’s ability to gather intelligence, formulate decisions, and execute air combat power rapidly and effectively against adversaries. The Air Force therefore has a need for a capability that rapidly detects DeepFake content across large volumes of PAI at scale, without being constrained by cryptologist and linguist skill and personnel shortages.  To address this need, 16th Air Force’s Air Force Cryptologic Office (AFCO) proposes to partner with Anonymous AI (d/b/a DeepMedia) on an SBIR Direct to Phase II (D2P2) to modify and adapt DeepID, DeepMedia’s commercial DeepFake detection and biometric analysis platform. DeepMedia will develop a customized deepfake detection platform so the AFCO can detect, attribute, and characterize DeepFake audio, video, and text footage. DeepMedia will perform non-recurring engineering (NRE) and research, development, testing, and evaluation (RDT&E) to develop a customized algorithm set and bespoke platform that enables AFCO analysts to assess open-source intelligence (OSINT), primarily through publicly available information (PAI) sources including social media, news, and press releases for DeepFake threats. Given AFCO’s mission set, the focus will primarily be on audio, voice, and language manipulation, as defined by the end-user. The end result will be the introduction of a custom, Air Force-specific DeepFake detection tool that empowers AFCO and broader information warfare and intelligence communities with a new capability to detect audio / language / voice media manipulation, attribute its source, and determine malicious intent faster and more efficiently.