Next-Generation Intrusion Prevention System Using Hardware-Accelerated Data Processing Units (DPUs)

To address the growing challenge of real-time DDoS mitigation in edge environments, this paper introduces a hardware-accelerated framework that combines deep learning with Data Processing Unit (DPU) acceleration. We propose a fine-tuned ResNet-18 model, adapted via transfer learning from ImageNet to network traffic analysis, and deploy it on a DPU using TensorFlow Lite quantization for energy-efficient, edge-compatible inference. The framework achieves 99.45% detection accuracy on the CICDDoS2019 dataset while mitigating 95 Gbps of malicious traffic, representing a 4.75× throughput improvement over CPU-based systems. By offloading processing to the DPU, our solution reduces CPU utilization to 25.5%, demonstrating how hardware acceleration improves real-time performance.