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

Advancements in convolutional neural network (CNN) have resulted in remarkable success in various computing fields. However, the need to protect data against external security attacks has become increasingly important because inference process in CNNs exploit sensitive data. Secure Memory is a hardware-based protection technique that can protect the sensitive data of CNNs. However, naively applying secure memory to a CNN application causes significant performance and energy overhead. Furthermore, ensuring secure memory becomes more difficult in environments that require area efficiency and low-power execution, such as the Internet of Things (IoT). In this paper, we investigated memory access patterns for CNN workloads and analyzed their effects on secure memory performance. According to our observations, most CNN workloads intensively write to narrow memory regions, which can cause a considerable number of counter overflows. On average, 87.6% of total writes occur in 6.8% of the allocated memory space; in the extreme case, 93.9% of total writes occur in 1.4% of the allocated memory space. Based on our observations, we propose an efficient integrity-tree structure called Countermark-tree that is suitable for CNN workloads. The proposed technique reduces overall energy consumption by 48%, shows a performance improvement of 11.2% compared to VAULT-128, and requires a similar integrity-tree size to VAULT-64, a state-of-the-art technique.