Core Concepts
A novel Processing-in-Memory (PIM) architecture with precision scaling techniques is proposed to efficiently detect malware by leveraging the computational capabilities of the memory subsystem.
Abstract
The paper proposes a novel approach for malware detection that utilizes a Processing-in-Memory (PIM) architecture. The key highlights are:
PIM Architecture:
The PIM architecture is designed to support compute-intensive applications like Convolutional Neural Networks (CNNs) for malware detection.
It consists of a hierarchical structure with DRAM clusters, each containing multiple LUT-based PIM cores connected via a router.
The PIM architecture can efficiently perform the mathematical operations required for CNN layers, such as convolution and max pooling, by programming the LUT cores.
Precision Scaling:
The paper employs uniform quantization to scale the precision of input data from 32-bit floating-point to 16-bit, 8-bit, and 4-bit integer types.
This precision scaling helps reduce the number of MAC operations, thereby decreasing the throughput and memory consumption without significantly impacting the malware detection accuracy.
Malware Detection Model:
Binary application files are converted into grayscale images and used to train CNN models like AlexNet, ResNet, VGG-16, and MobileNetV2 for malware detection.
The models trained on the 8-bit and 4-bit precision data achieve around 98% and 95% accuracy, respectively, demonstrating the effectiveness of the precision scaling approach.
Performance Evaluation:
The proposed PIM architecture achieves 4.02x higher throughput and 64.13x better energy efficiency compared to a state-of-the-art GPU (Pascal Titan X) for AlexNet inference.
It also outperforms other PIM architectures, such as DRISA and LAcc, in terms of both throughput (1.09x) and energy efficiency (1.5x).
The paper presents a comprehensive solution to the resource-intensive nature of malware detection model updates by leveraging the PIM architecture and precision scaling techniques, making it a promising approach for efficient and sustainable cybersecurity practices.
Stats
The paper presents the following key metrics:
Malware detection accuracy of 98% for 8-bit precision and 95% for 4-bit precision data.
Throughput improvement of 4.02x compared to a state-of-the-art GPU (Pascal Titan X).
Energy efficiency improvement of 64.13x compared to a state-of-the-art GPU (Pascal Titan X).
Throughput improvement of 1.09x and energy efficiency improvement of 1.5x compared to other PIM architectures.
Quotes
"The proposed PIM architecture exhibits a 1.09× higher throughput compared to existing Lookup Table (LUT)-based PIM architectures. Additionally, precision scaling combined with PIM enhances energy efficiency by 1.5× compared to full-precision operations, without sacrificing performance."
"The experimental results indicate that the proposed PIM is 74.62×, 64.13× more energy-efficient and has 4.02×, 45× higher throughput compared to the GPU and CPU respectively."