Hardware Trojan Attacks on the Reconfigurable Interconnections of Field-Programmable Gate Array-Based Convolutional...

Area of Science:

• Artificial Intelligence
• Computer Engineering
• Cybersecurity

Background:

• Convolutional Neural Networks (CNNs) are crucial for AI, with reconfigurable accelerators enhancing their performance.
• The security of these accelerators is a growing concern due to potential hardware Trojan attacks.
• Reconfigurable interconnection networks are vital components vulnerable to such threats.

Purpose of the Study:

• To propose a hardware Trojan targeting the reconfigurable interconnection network of FPGA-based CNN accelerators.
• To introduce a novel detection method using Physically Unclonable Functions (PUFs) to counter these attacks.
• To evaluate the effectiveness of the proposed detection technique.

Main Methods:

• A hardware Trojan was designed to alter data paths in the reconfigurable interconnection network, causing erroneous computations.
• Physically Unclonable Functions (PUFs), specifically arbiter-PUFs, were implemented for detection.
• Experiments were conducted on a Xilinx Zynq XC7Z100 platform using popular CNN architectures (LeNet, AlexNet, VGG).

Main Results:

• The hardware Trojan significantly degraded inference accuracy (8.93% to 86.20%) across tested CNNs.
• The PUF-based detection successfully identified the presence and location of hardware Trojans.
• The proposed solution incurred minimal hardware overhead (0.27%).

Conclusions:

• Reconfigurable CNN accelerators are vulnerable to hardware Trojan attacks targeting interconnection networks.
• PUF-based detection offers an effective and efficient solution for safeguarding these critical AI components.
• Addressing hardware security is essential for the reliable deployment of AI systems.