We are proposing to develop a distributed deep learning classifier that specifically addresses the technical challenges of accurate distributed classification for missile defense. We plan to provide detailed information to the final classification decision that allows engagement of adversary threats. It aims to utilize a distributed classification solution that trains the entire network in a single coherent manner versus baseline approaches where training occurs for local sensors and the fused output is then trained on how best to classify the fused solution. Pushing the training further upstream, creates an environment where the information from alternate sources is available to identify features that in and of themselves may not be important locally but provide a more robust understanding of the fused classification problem. We will address classification using data from physically distributed dissimilar sensors, an issue that is not emphasized in most multimodal deep learning research. We will demonstrate our approach using representative data from multiple sensors that have been used in counter UAS applications with the focus of demonstrating the technology. As we transition to Phase II, we will use more realistic data sources that represent the actual threat environment. Our work will compare multiple approaches for object classification to provide an understanding of the expected performance gain from our proposed approach. Approved for Public Release | 21-MDA-11013 (19 Nov 21)
