Experimental evaluation of algorithms for online network characterizations ONTIC: D4.3

National audienceDeliverable D4.3 aims at presenting the experimental evaluation of algorithms for online network characterization. These algorithms aim at characterizing the network by detecting anomalies in real time and in an unsupervised way.The first part of this document presents the experimental design exploited to test the platform and the algorithms, and provides detailed information about their configuration and parameterization. This deliverable is used as a base for the implementation of the use case 1 prototype in the context of WP5. Sections 4, 5 and 7 presents the works performed on unsupervised network anomaly detection. Section 4 describes the generation of a ground truth called SynthONTS for Synthetic Network Traffic Characterization of the ONTS Dataset. This ground truth is used to validate the unsupervised network anomaly detector presented in section 4. We claim that this ground truth is realistic, contains many different anomalies and is exhaustive in the anomaly labelling. Section 5 presents the unsupervised network anomaly detector proposed by ONTIC. It is an improved version of ORUNADA presented in deliverable D4.2. Section 7 describes the evaluation deployed to test the unsupervised network anomaly detector using the Google cloud platform and more specifically the Google Dataproc and the Google Storage.The second part of this deliverable addresses three different scenarios related to forecasting techniques and detection of anomalies. Section 8 describes our progress in network traffic behavior forecasting, as well as the obtained results when applied to the ONTS dataset. In particular, we show the application of deep convolutional neural networks in order to exploit the temporal nature of this forecasting scenario. Section 9 presents SLBN++ a proactive congestion control protocol equipped with forecasting capabilities that outperforms current proposals. Finally, Section 10 shows preliminary results of an approach to detecting anomalous behavior in cloud infrastructure based on deep neural networks