Automatic Secure Classification of Unmarked Sensor Data (ASCUS)

ABSTRACT: Intelligence, Surveillance, and Reconnaissance sensors deployed by the US Air Force generate enormous amounts of data, much of which may be used in a cross-domain security context. All of this data must be assigned an appropriate classification before it may be disseminated. Normally, classification information is stored as metadata attached to the data in question; unfortunately, there are a number of places where current classification processes fail to assign necessary security markings to sensor metadata, forcing human analysts to review the data manually to ensure its correct classification. We propose to address this shortcoming by designing and demonstrating a system supporting Automatic Secure Classification of Unmarked Sensor Data (ASCUS). ASCUS will process incoming sensor data, checking for empty fields and parsing out security-relevant concepts embedded in the metadata. This information will be passed into a reasoning engine, which will use probabilistic models derived from operationally relevant classification guides to determine a default classification level for the data. This default classification will be inserted into the data"s security metadata, along with an associated confidence level in the evaluated classification level. We will learn the parameters of our probabilistic models using machine learning techniques on an appropriate set of training data. BENEFIT: ASCUS"s ability to identify missing security metadata and automatically suggest default classification markings will ease the workload of intelligence analysts seeking to classify mistakenly unmarked sensor data, and help ensure mission-critical data is distributed in a secure and timely manner.