In-memory computing (IMC) systems have great potential for accelerating data-intensive tasks such as deep neural networks (DNNs). As DNN models are generally highly proprietary, the neural network architectures become valuable targets for attacks. In IMC systems, since the whole model is mapped on chip and weight memory read can be restricted, the pre-mapped DNN model acts as a “black box” for users. However, the localized and stationary weight and data patterns may subject IMC systems to other attacks. In this paper, we propose a side-channel attack methodology on IMC architectures. We show that it is possible to extract model architectural information from power trace measurements without any prior knowledge of the neural network. We first developed a simulation framework that can emulate the dynamic power traces of the IMC macros. We then performed side-channel leakage analysis to reverse engineer model information such as the stored layer type, layer sequence, output channel/feature size and convolution kernel size from power traces of the IMC macros. Based on the extracted information, full networks can potentially be reconstructed without any knowledge of the neural network. Finally, we discuss potential countermeasures for building IMC systems that offer resistance to these model extraction attack.
30 August 2022 in IEEE Journal on Exploratory Solid-State Computational Devices and Circuits
As more cloud computing resources are used for machine learning training and inference processes, privacy-preserving techniques that protect data from revealing at the cloud platforms attract increasing interest. Homomorphic encryption (HE) is one of the most promising techniques that enable privacy-preserving machine learning because HE allows data to be evaluated under encrypted forms. However, deep neural network (DNN) implementations using HE are orders of magnitude slower than plaintext implementations. The use of very long polynomials and associated number theoretic transform (NTT) operations for polynomial multiplications are the main bottlenecks of HE implementation for practical uses. This paper introduces RM-NTT, a resistive random-access memory (RRAM) based compute-in-memory (CIM) system to accelerate NTT and inverse NTT (INTT) operations. Instead of running FFT-like algorithms, RM-NTT uses a vector-matrix multiplication (VMM) approach to achieve maximal parallelism during NTT and INTT operations. To improve efficiency, RM-NTT stores modified forms of the twiddle factors in the RRAM arrays to process NTT/INTT in the same RRAM array and employs a Montgomery reduction algorithm to convert the VMM results. The proposed optimization methods allow RM-NTT to significantly reduce NTT operation latency compared with other NTT accelerators, including both CIM and non-CIM based designs. Effects of different RM-NTT design parameters and device non-idealities are also discussed.
Compute-in-Memory (CIM) implemented with Resistive-Random-Access-Memory (RRAM) crossbars is a promising approach for Deep Neural Network (DNN) acceleration. As the DNN size continues to grow, the finite on-chip weight storage has become a challenge for CIM implementations. Pruning can reduce network size, but unstructured pruning is not compatible with CIM, while structured pruning leads to higher neural network accuracy drop. In this work we systematically evaluate how structured pruning can be efficiently implemented in CIM systems. We show that by utilizing the inherent computational granularity in CIM operations, fine-grained structured pruning can be supported with improved accuracy and minimal hardware cost. We discuss the hardware implementation in a practical system and the expected performance in terms of accuracy, energy and effective throughput. With the proposed approach, compression ratio up to 11.1 (i.e. 9% weights remaining) can be achieved with only 0.6% accuracy drop with minimal hardware overhead in the hardware design.
In this letter, we demonstrate a physical unclonable function (PUF) system based on fingerprint-like random planar structures through pattern transfer of self-assembled binary polymer mixtures. With properly designed electrode structures and materials, different types of conductance distributions with large variations are achieved, allowing the PUF to function in differential or on/off modes. The uniqueness of the PUF is verified with inter Hamming distance and entropy at different temperatures. The PUF is shown to be resistant to machine learning attack based on Fourier extrapolation. The proposed fingerprint PUF is compatible with back-end-of-line (BEOL) process and offers potential for hardware security primitive in IoT industry.