Improving Managed Network Services Using Cooperative Synthetic Data Augmentation

Many managed service vendors in networking are adopting machine learning (ML) for many applications for their customers; e.g., anomaly detection, device fingerprinting, and resource management. Today, the data for training is siloed across customers leading to sub-optimal performance. While there are emerging proposals (e.g., federated learning, multi-party computation) to enable cooperative learning, these are at odds with analysts need for data for model exploration and testing. In this thesis, we envision a novel use of synthetic data generated using Generative Adversarial Networks (GANs) to augment the performance of existing ML workflows. We formulate the cooperative data augmentation problem, identify the design space of options, and identify key research challenges. We demonstrate the preliminary promise under two settings: (1) trac classification and (2) novelty detection showing that our improved workflow can enhance the performance of ML models up to 58% in AUC score. We also identify limitations and discuss for future work.