Machine-Assisted Declassification Review

The rapid growth of classified records across intelligence agencies, along with the public’s interest in declassification for democratic accountability, have created challenges for the Air Force Declassification Office (AFDO). We propose to research the nature of classified information, the objectivity and mechanizability of declassification criteria, and graded robust metrics for evaluating our approach. In Phase I, we will scope the classified subarea to be explored, blueprint a declassification architecture and explore data sources. We will also identify state-of-the-art tools and techniques that could be utilized to construct a prototype. This will provide the foundation for Phase II, where we expect to experiment with machine learning algorithms for declassification combining both data-driven and rule-based knowledge, and that incorporates selective human guidance for balancing efficiency with information security.