Deep Neural Network (DNN) is a data-hungry algorithm, which has a large energy cost on moving data between the memory and the computating unit. Recent works have proposed using ReRAM to design Process-In-Memory (PIM) accelerators to perform the computation inside the memory. However, the IP protection of the DNNs deployed on such accelerators is an important topic that has been less addressed. Firstly, due to its non-volatility, ReRAM does not need a continuous power supply to retain data. This makes the accelerator susceptible to new security vulnerabilities, for example, accessibility to the stored model if a device gets stolen. Secondly, because ReRAM's crossbar structure can only compute on cleartext data, encrypting the ReRAM content is no longer a feasible solution in this scenario. In this paper, we propose an IP protection solution on ReRAM based DNN accelerators to store DNN weights on crossbars in an encrypted format while still maintaining ReRAM's in-memory computing capability. The proposed solution stores and computes the DNNs in Stochastic Computing (SC) format, which can easily hide its conveyed weight values by scrambling the bit stream segments. However, SC's long bit streams incur a large storage overhead. To tackle this problem, we also propose two techniques to share bits among multiple weights, effectively compressed DNN's model size to reduce storage overhead.
Paper
Full text
A DNN Protection Solution for PIM accelerators with Model Compression
Semantic Scholar · Computer Science · 2022
Abstract
Deep Neural Network (DNN) is a data-hungry algorithm, which has a large energy cost on moving data between the memory and the computating unit. Recent works have proposed using ReRAM to design Process-In-Memory (PIM) accelerators to perform the computation inside the memory. However, the IP protection of the DNNs deployed on such accelerators is an important topic that has been less addressed. Firstly, due to its non-volatility, ReRAM does not need a continuous power supply to retain data. This makes the accelerator susceptible to new security vulnerabilities, for example, accessibility to the stored model if a device gets stolen. Secondly, because ReRAM's crossbar structure can only compute on cleartext data, encrypting the ReRAM content is no longer a feasible solution in this scenario. In this paper, we propose an IP protection solution on ReRAM based DNN accelerators to store DNN weights on crossbars in an encrypted format while still maintaining ReRAM's in-memory computing capability. The proposed solution stores and computes the DNNs in Stochastic Computing (SC) format, which can easily hide its conveyed weight values by scrambling the bit stream segments. However, SC's long bit streams incur a large storage overhead. To tackle this problem, we also propose two techniques to share bits among multiple weights, effectively compressed DNN's model size to reduce storage overhead.