(English only)

Bachelor's Thesis: Efficient Evaluation of the Fourier Transform of Multipoint Correlation Functions

This bachelor thesis deals with transforming Green's functions from the imaginary time domain to the frequency domain. Green's functions and their Fourier transform are useful entities, due to being directly measurable as physical quantities like the spectral function or the susceptibility. Complex many-body interactions can be modelled using two-, three- and multi-point functions. Due to the analytic evaluation in frequency space being very costly in comparison to evaluation in time, efficient Fourier transform methods are needed. Several numerical Fourier transform methods are compared with one another and assessed in terms of accuracy and runtime. The methods were applied to fermionic two-point, fermionic three-point and bosonic three-point correlation functions for Hubbard and Anderson impurity models. The main results of this work include the implementation of the exact representations of the correlation functions in frequency space, as well as several numerical methods for the Fourier transform. Especially for three-point functions, due to time-ordering and a resulting discontinuity, simpler approaches begin to fail, but the so-called SimplexQuad method yields more accurate and faster results than the other approaches by two orders of magnitude.

PDF: Efficient Evaluation of the Fourier Transform of Multipoint Correlation Functions

Master's Thesis: Explainability in Hate Speech Detection

This work examines the explainability of natural language processing on the example of hate speech and offensive content detection. After an introduction to the hate speech task is given, first a paper about our baseline systems on such a shared task is reviewed. Afterwards, the concepts of rationales and rational-based explainability metrics are presented, which are then used to compare not only the performance but also the explainability-metrics plausibility and faithfulness of three deep learning models with those of hand-made rule-based systems on the two tasks of detecting offensive text targeted against women and homosexuals. For these tasks, the dataset HateXplain is processed into smaller datasets specifically for detecting hate and offensive content against these specific target groups. Also, human annotations are evaluated in terms of their explainability for comparison. In the end, an qualitative error analysis is conducted.

We learn that rules perform better in precision and faithfulness and deep learning models in F1 score, some human-annotated rationales should not necessarily be viewed as gold labels and well-performing rules are not necessary rules which yield well explaining rationales. However, if the rules are engineered in a way to predict good rationales, explainability performance can be higher than deep learning models with and without attention-based mechanisms. Disclaimer: This work contains profane words.

PDF: Explainability in Hate Speech Detection

hateflux - Hatespeech and Offensive Content Detection Interactive Web Interface

hateflux - an AI for detecting hateful and offensive content! Due to my master thesis, I am developing several models to detect offensive content and hatespeech. I was now able to deploy such a reduced model to Microsoft Azure! Try it out yourself (Currently, English only): (Made with Milligram!)

Migration From HTTP to HTTPS and a new Domain

Websites with user input should definitely use HTTPS. Since 2014, Google uses HTTPS as a ranking signal. Therefore, even with static websites like the one from my uncle, a migration pays off. We took advantage of the situation to also switch to a more meaningful domain, and with the help of World4You, this was done in two days. After that, I was able to get the full old website running with HTTPS on Now, users are calmed down because Browsers display the site as secure, and nobody can spy on the customers looking for spare parts or repairs. Each legacy site can be migrated!

Bachelor's Thesis: Proof of Concept of Hacking Cryptocurrency Hardware Wallets

Securing cryptocurrency funds require long-term safety. A lot of hardware cryptocurrency wallets were developed, because they offer a way to store the private key offline. This thesis is split up in two parts. The first half analyzes five existing wallets in terms of their software and hardware security design and attestation methods. Then, it presents a classification of exiting attacks for these wallets.

The second half provides a proof of a work of a possible supply chain attack by emulating a hardware wallet on a USB device and tricking the victim into sending his funds to the attackers cryptocurrency account. It proves that wallet software puts too much trust in the hardware and that such a hardware wallet can be emulated. Therefore, the devices are missing important attestations.

PDF: Proof of Concept of Hacking Cryptocurrency Hardware Wallets
NEW Follow-up ACM paper: Better Keep Cash in Your Boots - Hardware Wallets are the New Single Point of Failure

Paper: Augmented Reality and its Upcoming Trends in Engineering

Currently, virtual reality and therefore also its related research fields experience a revival. This is mostly due to new hardware developments like the HTC Vive, Microsoft HoloLens and many others. Hence, also augmented reality is now getting a chance to develop further and make interesting breakthroughs. The goal of this paper is, firstly, to give a short overview of state of the art augmented reality technology in general. Then, its new chances and possibilities in the field of engineering are presented, with respect to the classic product lifecycle. In each phase of this lifecycle, possible ways to integrate augmented reality within are covered.

PDF: Augmented Reality and its Upcoming Trends in Engineering