A method and system for inferring a real time multimedia data stream to be of a given type among a group of types, comprising: training a machine learning model with a dataset of statistical data for respectively inbound and outbound streams of multimedia data streams, computing real time statistical data relative to each of the real time inbound stream and the real time outbound stream; and applying the trained machine learning model to the real time inbound stream and the real time outbound stream. The machine learning model comprises: a first layer of two parallel Convolutional Neural Networks, CNNs, for the statistical data of respectively the inbound (402-403) and the outbound (412-413) streams, a first of the two parallel CNNs (402, 412) being configured to detect features related to patterns of the computed statistical data, a second of the two parallel CNNs (403, 413) being configured to detect features related to average of the computed statistical data; and a second layer (404) of one CNN configured to detect features related to a combination of patterns and average of the statistical data for both the inbound and the outbound streams.