Processing-in-Memory Architecture with Precision-Scaling for Malware Detection

The wide adaptations of embedded systems in multiple fields have led to smart connectivity across devices and enhanced computation capabilities. Despite the vast applications in different areas, embedded systems face huge security threats. One of the critical security vulnerabilities is caused by malicious software a.k.a malware. Successful malware detection by employing Machine Learning (ML) is widely adopted in many systems. One of the prominent challenges in implementing neural network (NN) architectures is the requirement to have a large number of computational resources. Furthermore, the frequent movement of data between logic and memory units adds large overheads. Conversely, the IoT and edge devices are often limited in terms of the number of available resources. As a panacea, we introduce a PIM-based architecture to address such concerns and improve memory access latency. Such a paradigm further enriches the malware detection latency by mitigating the data transfer latency. To further improve the throughput and energy consumption, we employ precision scaling for the PIM-based malware detection in this work. We observe a malware detection accuracy of 98% with the proposed technique. Our proposed PIM architecture has $1.09 \times$ higher throughput than other traditional PIM architectures. Furthermore, precision scaling and PIM improve the energy efficiency by $1.5 \times$ compared to the full-precision operation without any penalty in performance.