Runtime Detection of Hardware Trojan in Untrusted Network-on-Chip

Traditionally, the computing system security has been tackled as a software-level problem. With the globalization of the modern semiconductor industry, the design of a system may involve different parties, especially for the design of system-on-chip (SoC) where the high complexity of the design has pushed designers towards using the third-party components that may have been modified with malicious additions without notice of the designer. Such a malicious modification (known as hardware Trojan) may be very small and can escape the scrutiny of the offline circuit-level testing and appear in the final product, which creates a new class of threats that have to be addressed at the hardware level and at the runtime. For SoCs, one of the most vulnerable components is the communication sub-system, called network-on-chip (NoC). NoC is able to access all components in the system and can be an easy target for varied attacks. This thesis aims at the hardware Trojan inside the router of a NoC. When activated, the Trojan can modify passing packets for two different purposes: 1) hijacking packets to harvest information carried by the packet; 2) manipulating the packet to alter the packet data. Both attacks can be tackled by the authentication. But unlike the software level authentication designs in the traditional networking system, the authentication in NoC has on-chip overhead issues. The thesis investigates three designs: one for packet hijacking attack detection, one for packet integrity attack detection, and one for Trojan location detection. Each design has a different dominant overhead issue. We, therefore, have a different optimization objective in the design. For the packet hijacking attack detection, we target the overhead on the chip cost and we present a customization design. For the packet data integrity authentication, the main concern is the bandwidth consumption caused by the large tag size and we introduce a progressive authentication scheme to reduce the bandwidth overhead. For the Trojan location detection, we propose a dynamic search algorithm to minimize the energy consumption. We have conducted a range of extensive experiments on each design, which demonstrate the effectiveness of our design approaches