TRUSTWORTHY EMBEDDED SYSTEMS BASED ON DEEP NEURAL NETWORKS

Author

Danushka Senarathna, Southern Illinois University CarbondaleFollow

Date of Award

5-1-2025

Degree Name

Doctor of Philosophy

Department

Electrical and Computer Engineering

First Advisor

Tragoudas, Spyros

Abstract

This dissertation focuses on the development of Deep Neural Networks (DNN)-based reliable embedded systems. The study has been conducted under three topics, spanning three application areas: power grids, computer vision, and Electronic Design Automation (EDA).The first study (Chapter 2) is to develop a novel reliable DNN-based approach embedded in the central controller of power systems to detect False Data Injection Attacks (FDIAs). The focus of this study was to assess the vulnerability of the existing time series analysis DNN-based methods and develop an enhanced method to alleviate failures in FDIA detection using state-of-the-art DNN architectures.Data drift detection in image classification DNNs is crucial for ensuring reliable operation. Existing methods do not identify the drift magnitude except for noise effects. The second topic of the dissertation (Chapter 3) focuses on developing a generalized method to detect and quantize data drift in image classification neural networks due to various effects such as noise and weather effects. The third topic (Chapter 4) focuses on developing a reliable DNN-based method to accelerate the evaluation of Boolean functions. The study assessed the learnability of Boolean functions by the DNNs and developed a novel Convolutional Neural Network (CNN) architecture to enhance accuracy. The speedup of the proposed DNN-based Boolean function evaluation approach was evaluated considering Graphics Processing Units (GPUs) and Memristor Crossbar Arrays (MCAs). Furthermore, the proposed method is applied in single Stuck-At-Fault (SAF) evaluation of combinational circuits and results are presented.