Efficient Integrity-Tree Structure for Convolutional Neural Networks through Frequent Counter Overflow Prevention in...

Area of Science:

• Computer Science
• Hardware Security
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are increasingly used, but their inference processes handle sensitive data, necessitating robust security measures.
• Hardware-based Secure Memory solutions exist but often incur substantial performance and energy overhead, especially for resource-constrained environments like the Internet of Things (IoT).

Purpose of the Study:

• To investigate memory access patterns in CNN workloads to understand their impact on Secure Memory performance.
• To propose an efficient integrity-tree structure optimized for CNN memory access characteristics.

Main Methods:

• Analyzed memory access patterns of CNN workloads, focusing on write-intensive regions.
• Developed and evaluated an efficient integrity-tree structure, Countermark-tree, tailored for CNNs.

Main Results:

• Identified that CNN workloads exhibit highly localized write patterns, with a significant percentage of writes concentrated in small memory regions.
• Countermark-tree demonstrated a 48% reduction in energy consumption and an 11.2% performance improvement over VAULT-128.
• The proposed Countermark-tree achieved comparable integrity-tree size to the state-of-the-art VAULT-64.

Conclusions:

• The localized memory access patterns in CNNs present unique challenges for traditional Secure Memory techniques.
• Countermark-tree offers an effective and efficient hardware-based solution for securing CNN data, particularly in IoT applications, by optimizing for these specific access patterns.