A collection of projects and theses created during my studies.
Theses
Bachelor Thesis: Continuously Learning Prediction of Pedestrian Movements at Intersections with Recurrent Neural Networks
Till Beemelmanns - RWTH Aachen University - 2017: This thesis presented a holistic approach for pedestrian trajectory prediction with recurrent neural networks at urban scenarios. State of the art Long Short-Term Memory designs where deployed in order to learn preprocessed trajectories in an end-to-end fashion. The architecture was tested and validated on publicly available pedestrian datasets. In addition to that, datasets obtained by laser scanner measurements of various road users and synthetic human trajectory datasets have been successfully applied. In all cases the neural network was able to predict future individual movements with a good accuracy for a given history trajectory. The overall prediction quality of the model was hereby measured by several performance evaluation criteria. On this basis, it was qualitatively shown that the proposed method outperformed two basic baseline models on all datasets. During the development of this work, a flexible Python framework was implemented that contains complex pre-processing and visualisation routines.
Download - Bachelor_Thesis_Continuously_Learning_Prediction_of_Pedestrian_Movements.pdf
Projects
Convolutional Neural Network and Recurrent Neural Network for Earthquake Detection and Localization
Till Beemelmanns - Michigan State University - Fall 2018 - 801 Computational Modeling: In order to improve seismic hazard assessment recent publications proposed machine learning and deep learning methods to detect and locate earthquakes. These approached are based on publicly available dataset that contain continuous waveform measurements and corresponding earthquake events. The most recent and most popular paper that applied deep learning to this kind of data was published by Perol et al. The researchers use Convolutional Neural Networks (CNNs) on waveform data to detect earthquakes and location. In the course of this project the results in using a CNN architecture have to be validated and in addition the prediction accuracy should be further improved. The application of Recurrent Neural Networks (RNN) using Long-Short Term Memory Cells (LSTM) were applied to this time-series classification problem, but proved to be ineffective. Finally, a fused architecture consisting of CNN layers and bi-directional LSTM cells was applied. A benchmark of both architectures on the same training and test datasets is performed.
Download - Project_Earthquake_Detection_and_Localization.pdf
Website Fingerprinting and Traffic Labeling with Deep Neural Networks
Till Beemelmanns & Adam Weckle - Michigan State University - Fall 2018 - 825 Computer Security: In our project we will investigate the capabilities of Deep Neural Networks for Website Fingerprinting and traffic labeling based on The Onion Router (Tor) network traffic. Users of the Tor network expect total anonymity while they are visiting websites and using online services. However, various researchers have proven that the network traffic that is exchanged between the user and the network can be used for Website Fingerprinting (WF) attacks. A possible eavesdropper between host and Tor entry node could capture encrypted routing information and the specific traffic pattern. The timing and the sizes of network packages can create a unique fingerprint for a specific website or set of websites. A possible attacker could use this pattern in order to determine which website Tor users are accessing. This method of attack was presented in various studies [1]–[5]. Based on these works, we want to improve the attack mechanism by using a different deep learning model and by predicting what type of file transaction the user is performing on these websites. This result is predicted to be more widely applicable to open-world applications.
Download - Project_Website_Fingerprinting_and_Labeling_with_Deep_Neural_Network.pdf
Github
Coming Soon