TUM Logo

Anomaly Detection Assisted by Generative Models

Anomaly Detection Assisted by Generative Models

Supervisor(s): Jan-Philipp Schulze, Philip Sperl
Status: open
Topic: Anomaly Detection
Type of Thesis: Masterthesis
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching


Master’s Thesis

Anomaly Detection
Assisted by Generative Models

Motivation and Task Description

In anomaly detection, we try to find data samples that deviate from our notion of normal. Usually, only a few anomalous examples exist, which is an especially challenging task for data-demanding machine learning frameworks like neural networks. Generative models, e.g. generative adversarial nets1, construct realistic new samples based on their training data. In our research2, we also found them useful for generating anomalous counterexamples.

Throughout your master’s thesis, you will research the state-of-the-art generative models and how they are integrated in current anomaly detection methods. Based on your knowledge, you will apply generative models to boost the detection performance of e.g. semi-supervised anomaly detection. Your main challenge is how to construct realistic anomalous counterexamples from normal data only, which generalise to yet unseen anomalies across multiple data types.


Sophisticated programming skills in Python First experience in deep learning libraries

Interest in machine learning and IT Security Motivation and self-organization


Jan-Philipp Schulze

Telefon: +4989322-9986-195

Philip Sperl

Telefon: +4989322-9986-141


Please attach your CV and your current transcript of records to your application.

Fraunhofer Research Institution for Applied and Integrated Security (AISEC) Cognitive Security Technologies (CST)
Lichtenbergstraße 11, 85748 Garching (near Munich), Germany

1Generative Adversarial Nets; Goodfellow et al.; NIPS 2014; http://papers.nips.cc/paper/ 5423-generative-adversarial-nets

2Activation Anomaly Analysis; Sperl, Schulze, Bo ̈ttinger; ECML-PKDD 2020; https://arxiv.org/ abs/2003.01801