MRMafizur Rahman

Mafizur Rahman

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Mafizur Rahman has a total of 3 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Comparative Analysis of Inference Performance of Pre-Trained Convolutional Neural Networks in Analog Accelerators

    by Mafizur Rahman, Lijun Qian

    Resistive crossbars using non-volatile memory devices have become promising components for implementing deep learning models in hardware. However, crossbar-based computations encounter a notable challenge due to device and circuit-level non-idealities. In this study, we explore three recent crossbar simulation tools, namely, AIHWKIT, CrossSim, and MemTorch, and evaluate four pre-trained convolutional neural networks (CNNs) for image classification on CIFAR-100, MNIST, and SVHN datasets using these tools. To the best of our knowledge, this is the first study using all three simulation tools of analog accelerators, representing three distinctive hardware settings, to evaluate CNN-based model robustness under analog noise and nonlinearities. We first test the robustness of four models (VGG19, InceptionV3, ResNet50, & MobilenetV2) using different levels of white Gaussian noise as baselines. Then we evaluate their inference performance on the three tools to determine their resilience to analog noise and nonlinearities in the hardware environment. Results show that while all CNNs' suffer performance degradation as expected, insights are obtained such as ResNet50 outperforms others in two out of three simulators despite real-world hardware imperfections due to its deep structure. Furthermore, we explore the impact of three key factors: cell bits, ADC resolution, and tile size on CNN performance across various hardware configurations, providing recommendations for the efficient deployment of CNN on hardware. Our results suggest that increasing cell bits and ADC resolution improves CNNs' performance. It is also observed that smaller tiles are suitable for lightweight models, while for complex networks, we should choose larger shape of tiles.


  • Open Access

    Evaluating Pretrained Deep Learning Models for Image Classification Against Individual and Ensemble Adversarial Attacks

    by Mafizur Rahman, Prosenjit Roy, Sherri S. Frizell, Lijun Qian

    The robustness of Deep Neural Networks (DNNs) against adversarial attacks is an important topic in the area of deep learning. To fully investigate the robustness of DNNs, this study examines four frequently used white box adversarial attack techniques, namely, the Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), Basic Iterative Method (BIM), DeepFool, and their effects on DNN models for the image classification task. The results show that ResNet152 and DenseNet201 are less vulnerable comparing to other DNN models to a variety of individual attacks, highlighting their intrinsic strength even in the lack of specific adversarial training. Further, we propose two ensemble adversarial attacks combining three individual attacks for generating adversarial examples from the tiny ImageNet, CIFAR-10, CIFAR-100, and SVHN datasets for DNN model evaluation. It is observed that the performance of the DNNs deteriorate significantly under the proposed ensemble adversarial attacks even after defensive measures have been applied. For instance, the accuracy of the most robust DNN that we tested, namely the defense distillation enhanced DenseNet201, dropped more than 59% under the proposed ensemble adversarial attacks, comparing to only 34% decrease under the individual attacks.