Preparing Deep Learning Models with Minimal Training Data Using Adversarial Domain Adaptation

Abstract Most military deployments comprise of a plethora of interconnected devices, personnel and network equipment. These deployments might have their own deep learning (DL) models for assisting with various analytical and decision-making tasks like network intrusion detection, resource sharing, access control, human activity recognition etc. Training DL models, however, requires vast amounts of labeled data which is often expensive and time-consuming to collect. Due to the dynamic nature of these deployments, the DL models need to be updated regularly to adapt to the new situations and contexts. Also, DL models trained using data from one deployment may not be applicable to other deployments due to differences in their individual composition. To address the problem of scarcity of labeled training data in a domain, we propose the use of adversarial domain adaptation. This approach allows us to transfer knowledge gained from learning on an existing domain to a related domain with very less labeled data. In this demonstration, we show that our proposed approach can create highly accurate DL models for network intrusion detection, even when the number of labeled samples in the target dataset is significantly small
  • Ankush Singla (Purdue)
  • Seraphin Calo (IBM US)
  • Elisa Bertino (Purdue)
  • Dinesh Verma (IBM US)
Date Sep-2020
Venue 4th Annual Fall Meeting of the DAIS ITA, 2020