Applying AI for and on Hardware: Cross-Layer Approaches for Security and Efficiency

With the emergence of several powerful new models, AI became crucial in hardware design. It enhances performance, efficiency, and security. This paper examines AI in hardware security and optimization. We introduce AI methods for securing hardware: pre-deployment, at runtime, and post-deployment. Additionally, we present AI-driven strategies for system-level resource management on emerging manycore platforms with 3D-stacked memory, deploying neural networks to predict performance and thermal dynamics under migration and DVFS policies. On the other hand, we also improve the execution of ML models on hardware. First, we propose a training-time regularization technique for decision trees and random forests that enforces asymmetric splits, improving cache behavior with minimal accuracy loss. Second, we evaluate multiple model compression techniques for storing decision trees in memory-efficient formats, showing that specific schemes can simultaneously reduce memory and improve inference latency. Our methods, tested on real hardware or validated datasets, highlight trade-offs in model complexity, hardware constraints, and system-level performance. We support AI and hardware co-design for adaptive, efficient, and trustworthy systems.