Efficient Neural Computation on Network Processors for IoT Protocol Classification

The Internet of Things (IoT) brings forth pressingrequirements on the service providers in terms of service differen-tiation, which plays an important role in pricing policies as well asnetwork load balancing. In this paper, we consider differentiationof application level protocols for IoT from general applicationprotocols through flow classification. We implement a neuralnetwork classifier that can run at wire speed reaching 100 Gbpson a network processor. In particular, we study approximationswhich allow us to efficiently compute the neural network output,while complying with the network processor limitations, whichdoes not provide multiplication or other complex mathematicaloperations. The results show that the implementation is efficientand that the classification error is negligible.