University of Southampton
Deep learning enhanced security for network-on-chip in heterogeneous computing systems
Abstract
dc:description.abstractThe rapid advancement of Artificial Intelligence (AI) has accelerated significantly in recent decades. Notable milestones include ImageNet, which established a benchmark for vision-based AI tasks, and Google DeepMind's AlphaGo defeating human players. Currently, Large Language Models (LLMs) such as ChatGPT and Deepseek are driving the ongoing AI revolution. Correspondingly, specialized AI chips have emerged to optimize model execution, improve energy efficiency, and accelerate AI development, creating substantial commercial opportunities. However, the increasing complexity and widespread adoption of these chips have raised significant security concerns.<br/><br/>Deep Learning (DL), a cornerstone of modern AI, requires computing platforms capable of efficiently, securely, and reliably handling intensive computation and communication tasks. Due to DL’s need for both parallel and sequential processing, heterogeneous computing architectures have been adopted by leading fabless chip designers, including Nvidia, AMD, Intel, and Apple Silicon. This has made robust on-chip communication systems essential. Network-on-Chip (NoC) technology, commonly utilized in heterogeneous Multi-Processor Systems-on-Chip (MPSoC), plays a critical role in managing data exchange efficiently, similar to blood circulation within an organism.<br/><br/>However, NoCs are vulnerable to security threats such as Data Packet Tampering (DPT) and Flooding Denial-of-Service (FDoS) attacks, severely affecting system performance and communication integrity. This thesis introduces novel threat models and countermeasures, including an Artificial Neural Network (ANN)-based method for detecting and localizing DPT attacks and a Convolutional Neural Network (CNN)-based model to accurately detect and localize FDoS attacks with low hardware overhead. The research also explores the combined use of Graph Neural Networks (GNN) and Long Short-Term Memory (LSTM) models for early detection and localization of FDoS attacks. These proposed detection mechanisms have been benchmarked and validated using relevant workloads, demonstrating their effectiveness.<br/><br/>Investigating these NoC security threats and developing detection methods naturally leads to considering the security of the broader heterogeneous system, including the Deep Learning accelerators themselves. Therefore, in the final stage, this thesis integrates the FDoS attack model into a heterogeneous DL accelerator modeling platform featuring NoC, Central Processing Unit (CPU), Direct Memory Access (DMA), and various other system components. Eight identified security threats, covering runtime on-chip communication security and runtime accelerator security, have been modeled and implemented in this thesis for this platform. This comprehensive security threat modeling platform offers valuable insights for future research and practical development in secure AI accelerator hardware.<br/><br/>Overall, this thesis provides a thorough introduction to the state-of-the-art in NoC security, presents detailed threat modeling and innovative countermeasures leveraging Deep Learning, and introduces a versatile platform for security research in heterogeneous DL accelerators. These contributions highlight the significant potential for employing DL techniques to enhance hardware security. As a step towards realizing this potential architecturally and bridging the gap between detection algorithms and hardware integration, this research also explored the conceptual design of STAR-NoC (Security-Targeted Router-based Network-on-Chip). This microarchitecture concept proposes embedding deep learning-based detection and localization modules directly into the NoC fabric, envisioning an efficient, scalable, and resilient approach for integrating security functions within the communication infrastructure. Thus, the work not only develops specific security solutions but also lays groundwork for future secure hardware architectures.
Degree
thesis:*- Name dc:type.qualificationname
- Ph.D.
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- University of Southampton
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Wang, Haoyu
- Advisors dc:contributor.advisor
-
- Halak, Basel
- Atamli, Ahmad
- ORCID person:identifier.orcid
Rights
- Language dc:language
- English, eng