Title:

Ensemble Methods für Deep Learning based Anomaly Detection

Description:

This work explores a modified bagging approach utilizing deep learning base models: Sampling without replacement, fitting of an autoencoder on each sampled subset and maximum-aggregation of the reconstruction errors of each base model into one anomaly score. The performance of the proposed ensemble method is compared to a range of classical anomaly detection models, single autoencoders as well as to the regular bagging ensemble and evaluated on datasets originating from different domains. Effectiveness of the proposed method was shown for the MNIST dataset, while the ensemble method did not provide any improvements compared to baseline models on the Cifar-10 and HDFS log dataset.

Supervisor: Boris Lorbeer

Type:  Master Thesis

Duration: 6 months