Robust Learning-Enabled Intelligence for the Internet of Things: A Survey From the Perspectives of Noisy Data and Adversarial Examples

Abstract

The Internet of Things (IoT) has been widely adopted in a range of verticals, e.g., automation, health, energy, and manufacturing. Many of the applications in these sectors, such as self-driving cars and remote surgery, are critical and high stakes applications, calling for advanced machine learning (ML) models for data analytics. Essentially, the training and testing data that are collected by massive IoT devices may contain noise (e.g., abnormal data, incorrect labels, and incomplete information) and adversarial examples. This requires high robustness of ML models to make reliable decisions for IoT applications. The research of robust ML has received tremendous attention from both academia and industry in recent years. This article will investigate the state of the art and representative works of robust ML models that can enable high resilience and reliability of IoT intelligence. Two aspects of robustness will be focused on, i.e., when the training data of ML models contain noises and adversarial examples, which may typically happen in many real-world IoT scenarios. In addition, the reliability of both neural networks and reinforcement learning framework will be investigated. Both of these two ML paradigms have been widely used in handling data in IoT scenarios. The potential research challenges and open issues will be discussed to provide future research directions.