DARPA’s Semantic Forensics program reaches final stage, promising commercialization of deepfake defense

As a result of the Semantic Forensics (SemaFor) program, DARPA recently reported that defense technology to help against manipulated media may be ready for commercialization.

“Our investments have seeded an opportunity space that is timely, necessary, and poised to grow,” Dr. Wil Corvey, DARPA’s Semantic Forensics program manager, said. “With the help of industry and academia around the world, the Semantic Forensics program is ready to share what we’ve started to bolster the ecosystem needed to defend authenticity in a digital world.”

Coming at a time of heightened scrutiny in the midst of another presidential campaign season, DARPA noted that its investments have systematically reduced developmental risks and made it easier than ever for companies and academia alike to research the subject and capitalize on the defensive groundwork. As a result, it intends to launch two new endeavors: an open source catalog and a community research effort.

The analytic catalog will contain open-source resources developed under SemaFor for use by researchers and industry, with more being added all the time. Known as the AI Forensics Open Research Challenge Evaluation (AI FORCE), the community research effort will, on the other hand, push developments of new machine learning models that can accurately deduce synthetic AI-generated images and otherwise manipulated or edited pieces. That effort goes live this week.

Despite all this, though, DARPA warned that a concerted effort will be needed across commercial, media, research and development and governmental interests to develop and utilize means of countering manipulated media. SemaFor merely exists to create the tools, but DARPA made clear it’s up to others to use them. That may be more critical than ever, as recent years have seen automated manipulation technologies become more accessible and, coupled with social media, provided an easy environment to help manipulated images go viral.