Efficient Neural Computation on Network Processors for IoT Protocol Classification

Pant, Vibha; Welponer, Michele; Rizzon, Luca; Lavagnolo, Roberto

doi:10.1109/ngcas.2017.55

Cited by 3 publications

(1 citation statement)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Among the methods that are considered, there are unsupervised learning algorithms, which directly create classification, such as k-means, Expectation Maximization (EM) and Density-Based Spatial Clustering of Applications with Noise; and supervised methods, which require a learning phase with ground truth, such as J48, Naive Bayesian (NB) and Support Vector Machine (SVM). In our previous work, we have exploited neural networks as a supervised learning method [23]. Two interesting results can be learned from these studies.…”

Section: Related Workmentioning

confidence: 99%

Packet Length Spectral Analysis for IoT Flow Classification Using Ensemble Learning

Cirillo

2020

IEEE Access

Self Cite

View full text Add to dashboard Cite

With the proliferation of ubiquitous and autonomous devices for sensing, control, monitoring and conditioning, the Internet of Things (IoT) holds a great potential for the development of innovative applications. At the same time, network operators must support these devices with differentiated services, which rely on the ability to automatically recognize and classify the nature of the communication flows. In this paper, we present a supervised learning approach to discriminate between IoT and non-IoT traffic, and to determine the class of the device originating the packet flows. We make use of a reduced set of features based on the spectral analysis of the packet lengths of a flow, and evaluate an ensemble learning algorithm that uses a Random Forest classifier. We first discuss the datasets and the procedure that we use to extract the features, with examples from different devices. The evaluation is performed using both 10-fold cross validation and a split between training, validation and test-set. The latter is used for hyperparameter tuning. The results show that for reasonably large datasets the classifier achieves very high accuracy, as well as Precision and Recall rates. We further improve the performance on individual devices by selectively replicating the flows in the dataset, to achieve a better balance. We then evaluate a real-time implementation, and propose a runtime procedure to evaluate the model confidence level and trigger a retraining phase to adapt to a changing environment. A detailed analysis of the performance shows that the algorithm can support networks up to 100 Gbps on standard computing platforms.

show abstract

Section: Related Workmentioning

confidence: 99%