Finding anomalies in software licensing logs using unsupervised methods

Abstract: Cryptolens is world leading software licensing platform. As a result, it has large amounts of data that is generated when each end user application at- tempts to verify a license key. Being able to differentiate between normal and anomalous data can provide software vendors with a way to detect fraud and other abnormal behaviour, allowing them to save time on analyzing all the data themselves and increase revenues. It is found that an effective way to find anomalies in software licensing logs is to use the reconstruction error as the anomaly score from either an LSTM or TCN based autoencoder, where the decision boundary is decided by the largest error in the error histogram on the training set.