ACD3: Acoustic Context Data Drift Detector

The US Navy (USN) is facing increasingly competent competitors at sea, including pacing threats with submarine fleets (such as the Type 094 or anticipated Type 096 SSBN), making detect-to-engage difficult and/or delayed. To detect the subsurface threats, USN surface vessels employ passive sonar to detect, classify, and track contacts. Artificial Intelligence and Machine Learning (AI/ML) algorithms show great potential in improving accuracy and speed of submarine detection and tracking. Constructing these algorithms is costly and time-consuming, and updating them to include new data also requires significant effort. Expert human analysts must evaluate and label large quantities of data, then identify which data should be used in the retraining process. Determining which data samples to keep and reject in a time efficient manner is key to accurate models and fast redeployment. Fast redeployment defends USN vessels from evolving adversarial platforms and techniques, tactics, and procedures (TTPs).   Boston Fusion Corp (BFC) proposes the Acoustic Conditioned Data Drift Detector (ACD3), a novel self-supervised learning (SSL) technique that will enable the Navy to train AI/ML algorithms more efficiently, starting with acoustic data from undersea warfare systems. This technique uses conditioned latent reconstruction as a metric for data priority on new data or existing datasets. ACD3 minimizes the time spent by expert human analysts and accelerates model re-deployment by distinguishing between valuable diverse/edge case data (dataset expansion), redundant data (valuable for dataset balance), and out of scope data. ACD3 does not need ground truth (GT) labels during inference and is generalizable to other domains and sensor modalities making it relevant to other Navy workflows. ACD3 is easy to integrate into existing model redeployment pipelines to reduce the size of a dataset (decreasing labeling and training time).