Robust solutions for data storage and learning systems

Abstract

Within the last decade, machine learning (ML) has become the cornerstone of modern industries, driving a broad range of innovations. This thesis focuses on the “Learning from data” pipeline which supports such innovation. This pipeline consists of two main building blocks: 1) data storage/management system and 2) learning algorithms that extract key features from the collected data. Unfortunately, the performances of these building blocks are degraded by two practical issues: inevitable failure events and adversarial attacks. Providing reliability and robustness of “Learning from data” framework against such practical issues is the main objective of this dissertation. In the first part, we suggest clustered distributed storage system (C-DSS) which reflects the clustered nature of cloud storage, and propose theoretically optimal coding techniques for enabling the reliability of C-DSS against node failures, by leveraging tools from coding and information theory. In the second part, we suggest a framework called Election Coding, which guarantees the robustness of communication efficient distributed learning algorithms against Byzantine attacks. In the third part, we propose a data augmentation scheme called GAN-mixup, which improves the robustness of learning systems against adversarial attacks, by making use of conditional GANs to learn the class-conditional manifolds and generate data points in-between the manifolds of different data classes.