SquareLoop: Explore Optimal Authentication Block Strategy for ML

Abstract

Off-chip memory in ML accelerators is vulnerable to both hardware and software attack, which needs encryption and authentication. Precise performance modeling of it requires (1) representation of authentication blocks (AuthBlock) to cover the full design space of shapes and orientations, and (2) precise memory behavior modeling, as encryption and authentication mainly increase memory traffic. This paper introduces 𝑆 2Loop, a framework that resolves these challenges by introducing (1) flexible, all-level partitioning based AuthBlocks for ensuring full coverage of the entire design space, (2) a realistic layout-based memory model, and (3) an Mapping-LayoutAuthentication co-search algorithm to explore the drastic combinatorial design space to figure out optimal mapping, layout, and AuthBlock shape choice for multi-layer workloads. SquareLoop’s detailed memory model helps find better mapping to achieve 1.32× speedup on ResNet18 compared to the SotA SecureLoop, and our latency predictions are validated to within 7.3% of an RTL implementation. 𝑆 2𝐿𝑜𝑜𝑝 also achieve up-to 1.08×/1.82× overall speedup for authenticated ResNet18/MobileNet-V3 on various accelerators with AuthBlock and Mapping co-searching. We open-source 𝑆 2Loop to provide a powerful and validated tool for designing efficient, secure accelerators at https://github.com/maeri-project/squareloop.