British Telecommunications Public Limited Company

Yipeng Cheng

Maximilien Servajean

Giulio Giaconi

A computer implemented method of detecting anomalous behavior within a computer network, the method including accessing data records each corresponding to an occurrence of communication occurring via the computer network and including a plurality of attributes of the communication; generating, for each of at least a subset of the data records, a training data item for a neural network, the training data item being derived from at least a portion of the attributes of the record and the neural network having input units and output units corresponding to items in a corpus of attribute values for communications occurring via the network; augmenting the training data by replicating each of one or more training data items responsive to one or more attributes of the data record corresponding to the training data item; training the neural network using the augmented training data so as to define a vector representation for each attribute value in the corpus based on weights in the neural network for an input unit corresponding to the attribute value; repeating the accessing, the generating, the augmenting and the training to generate multiple generations of vector representations for each attribute value in the corpus, each generation corresponding to data records received during a different time period; and for at least a subset of attribute values in the corpus, comparing the multiple generations of vector representations to identify a change in one or more vector representation as an indication of an anomalous change of behavior in the computer network.

Detecting anomalies in computer networks

A method of anomaly detection for network traffic communicated by devices via a computer network, the method including clustering a set of time series, each time series including a plurality of time windows of data corresponding to network communication characteristics for a device; training an autoencoder for each cluster based on time series in the cluster; generating a set of reconstruction errors for each autoencoder based on testing the autoencoder with data from time windows of at least a subset of the time series; generating a probabilistic model of reconstruction errors for each autoencoder; and generating an aggregation of the probabilistic models for, in use, detecting reconstruction errors for a time series of data corresponding to network communication characteristics for a device as anomalous.

Anomaly detection in computer networks

A method of anomaly detection for network traffic communicated by devices via a computer network, the method including receiving a set of training time series each including a plurality of time windows of data corresponding to network communication characteristics for a first device; training an autoencoder for a first cluster based on a time series in the first cluster, wherein a state of the autoencoder is periodically recorded after a predetermined fixed number of training examples to define a set of trained autoencoders for the first cluster; receiving a new time series including a plurality of time windows of data corresponding to network communication characteristics for the first device; for each time window of the new time series, generating a vector of reconstruction errors for the first device for each autoencoder based on testing the autoencoder with data from the time window; and evaluating a derivative of each vector; training a machine learning model based on the derivatives so as to define a filter for identifying subsequent time series for a second device being absent anomalous communication.

Growing community of inventors

London, United Kingdom

Yipeng Cheng