15 January 2024 in IEEE Transactions on Information Forensics and Security
The Physical Unclonable Function (PUF) is a security mechanism that takes advantage of the physical variations in a device to create a unique response that can be used as a device signature or secure key. However, many DRAM-based PUFs violate the operating rules of commodity DRAM to exploit a source of entropy in the DRAM read path. This work proposes a fast and reliable DRAM-based PUF that evaluates the variation of adjacent cells and produces the response through the normal read operation. The proposed design is implemented using 65-nm technology, and a detailed statistical SPICE simulation verifies its validity. The statistical analysis shows that the proposed PUF achieves 54.19% uniformity and 49.43% uniqueness, with 98% of the investigated responses achieving a Shannon entropy of 0.95. Additionally, the proposed design generates the response by 45, which is at least 66.7 times faster than existing systems. Furthermore, the proposed design uses the relative behavior of cells, which allows for stable responses against temperature and voltage variations, eliminating the need for error correction codes. The proposed PUF also shows resiliency against machine learning-based modeling attacks, as the prediction accuracy does not exceed 55% over 5K Challenge-Response Pairs (CRPs). The area overhead is negligible as the proposed design uses standard circuits, with the addition of only one 2x1 multiplexer at the inputs of the row buffer.
18 March 2025 in IEEE Transactions on Circuits and Systems for Artificial Intelligence
Achieving energy efficiency is of paramount importance in facilitating the deployment of artificial intelligence (AI) applications, particularly for edge devices. The computational complexity and memory requirement of AI workloads have made energy efficiency in circuits the primary design constraint, prompting a need for architectural reconsideration. However, operations in Convolutional Neural Networks (CNN) relying on matrix multiplication offer opportunities for algorithm-level optimization. The dynamic power consumption in CMOS technology exhibits a direct correlation with the switching activities (SA) associated with transitions of data bits from one cycle to the next. Minimizing SA constitutes an effective approach to curbing power consumption, resulting in prolonged battery life. This work introduces a cross-layer end-to-end optimization technique, which involves a one-time sorting of pre-trained model parameters to reduce switching activity during matrix multiplication or convolution operations while eliminating the indexing overhead during inference. While upto 40% reduction in switching activities through sorting has been achieved, the actual saving in dynamic power achievable for the input stationary systolic array is approximately 10% in total power consumption and around 15% in switching power. The proposed method has been validated using various pre-trained networks, including GoogeLeNet, MobileNet, AlexNet, and SqueezeNet, among others.