Efficient Neural Computation on Network Processors for IoT Protocol Classification

The Internet of Things (IoT) brings forth pressing requirements on the service providers in terms of service differentiation, which plays an important role in pricing policies as well as network load balancing. In this paper, we consider differentiation of application level protocols for IoT from general application protocols through flow classification. We implement a neural network classifier that can run at wire speed reaching 100 Gbps on a network processor. In particular, we study approximations which allow us to efficiently compute the neural network output, while complying with the network processor limitations, which does not provide multiplication or other complex mathematical operations. The results show that the implementation is efficient and that the classification error is negligible.