Re-architecting Traffic Analysis with Neural Network Interface Cards

Explore a groundbreaking approach to enhance the scalability of online machine learning-based network traffic analysis in this 14-minute conference talk from USENIX NSDI '22. Discover how replacing traditional supervised machine learning models with binary neural networks can revolutionize traffic analysis. Learn about Neural Networks on the NIC (N3IC), an innovative system that compiles binary neural network models for direct integration into SmartNIC data planes. Examine the implementation and evaluation of this solution through two use cases: traffic identification and anomaly detection. Understand how N3IC achieves up to 100x lower classification latency and 1.5-7x higher throughput compared to state-of-the-art software-based systems. Delve into the design and FPGA-based prototype of a hardware primitive that adds binary neural network support to NIC data planes, requiring minimal logic and memory resources. Gain insights into how this technology enables more challenging use cases requiring microsecond-level online traffic analysis.