Preliminary Application of Neural Networks to Support Assessment of Ground-Based Imagery for International Safeguards Analysis

The International Atomic Energy Agency’s Department of Safeguards collects and analyzes information from open sources to support its all-source evaluation of safeguards-relevant data as part of the state evaluation process. Currently, collection and assessment of ground-based imagery is conducted manually using ad hoc (analyzing relevant images discovered alongside other open source data) or targeted (intentional search and collection of images related to a location) approaches. The proliferation of photo sharing via social media platforms has dramatically increased the availability of geo-tagged, ground-based imagery that may be relevant for safeguards analysis, but manual collection and analysis of this data would be impractical. Recent advances in data science enable the application of novel analysis and modeling techniques to support preliminary assessment of data sources such as ground-based imagery that can reduce the volume of irrelevant data and prioritize images for expert analyst assessment. One such advance is convolutional neural networks (CNNs), a form of machine learning algorithm loosely inspired by the architecture of the visual cortex, which allows computers to learn by example, and achieves state-of-the-art performance in image classification and related computer vision tasks. In this paper, we describe our work using CNNs with ground-based imagery for safeguards-relevant applications, along with preliminary results and a discussion of future work.