IMPACTS OF ADVERSARIAL MACHINE LEARNING METHODS IN DEEP LEARNING MODELS USED IN IOT ENVIRONMENTS

Abstract

The Internet of Things (IoT) and Artificial Intelligence (AI) have revolutionized various industries by enabling the creation of intelligent systems that can collect and analyze large amounts of data from connected devices and generate insights to support decision-making processes. However, the intersection of IoT and AI has also brought novel challenges, particularly in the realms of security and data privacy. The objective of this thesis is to demonstrate the impact that adversarial machine learning techniques can have in deep learning models used in IoT environments. We make three main contributions in this thesis. Firstly, we develop a white-box adversarial attack methodology and demonstrate its severity in degrading the targeted classifiers' performance. Secondly, we develop a black-box adversarial attack method which is designed to evade an entire class of defence methods, while significantly degrading the targeted classifiers' performance. Finally, we leverage these adversarial methods in a game-theoretic setting to develop a synthetic data generation methodology. These synthetic datapoints are used to improve the model performance, and retrain the classifiers without using any real-world data collected from end users, thus mitigating the privacy concerns that end-users may have regarding their personal data being collected by IoT applications.