Fast Malicious Packets Inspection Framework Using Converged Accelerator

The exponential growth in Internet data and the increasing sophistication and frequency of network attacks have intensified concerns over network security. This study introduces a packet inspection framework on DPU-GPU converged accelerators, leveraging GPU capabilities to enhance both the accuracy and efficiency of deep learning (DL)-based malicious packet detection. The offloading of DL model processing to the GPU significantly reduces the CPU workload, thus enhancing overall system efficiency. Nevertheless, we observed that, during periods of high network traffic, limitations in the accelerator's throughput can create bottlenecks. To address this limitation, we implemented an adaptive packet sampling mechanism that prevents system overload and maximizes the detection of malicious packets. Experimental results demonstrate that the proposed framework effectively balances detection accuracy and system performance, underscoring the efficacy of converged accelerators in real-time network security applications.