eprints (HSR Hochschule für Technik Rapperswil)
Not a member yet
1196 research outputs found
Sort by
Enhancing Cybersecurity with Machine Learning: Beaconing Detection in PCAP Data
No full textEnhancing Cybersecurity with Machine Learning: Beaconing Detection in PCAP Data
Abstract
This study explores the enhancement of cybersecurity through the application of machine learning techniques, specifically focusing on the detection of beaconing activity in network traffic (PCAP) data. PCAP, or packet capture, refers to the process of intercepting and logging traffic that passes over a computer network.
Beaconing, a communication technique and a common indicator of malicious activity requires complex multilevel detection methods due to its discreet and repetitive nature. My approach involves the development of a dual-model framework with a combination of a Histogram Gradient Booster Classifier (HGBC) and a Long Short-Term Memory (LSTM) neural network. The HGBC classifies the initial features extracted from the PCAP data, while the LSTM model further refines the detection by capturing temporal dependencies between consecutive packet flows.
The combined model achieves an accuracy rate of 99.37%, demonstrating its effectiveness in identifying beaconing patterns. This high level of accuracy illustrates the potential of a combination of machine learning and deep learning algorithms in advancing cybersecurity measures for unmasking threats in network traffic analysis
Analyse und menschzentrierte Erweiterung von lost.university
No full textDas Zusammenstellen eines eigenen Curriculums ist ein zentraler Bestandteil des Informatikstudiums an der OST - Ostschweizer Fachhochschule und kann für Studierende eine
Herausforderung darstellen. Diese Studienarbeit fokussiert sich auf die menschzentrierte Weiterentwicklung der Plattform lost.university, einem digitalen Werkzeug zur Unterstützung von Informatikstudierenden bei der Planung ihres Studiums an der OST. Ziel der Arbeit war die Entwicklung eines durch Usability-Tests validierten Prototyps, der die Benutzerfreundlichkeit sowie die Funktionalität der Plattform verbessert und den spezifischen Anforderungen der Zielgruppe gerecht wird.
Methodik
Als methodischer Rahmen diente ein Vorgehensmodell, welches stark an den User-Centered-Design- Prozess anlehnt. Zentrale Grundlage der Entwicklung war die Verstehen-Phase, in der
Anforderungen und Erkenntnisse aus den im Synthese-Workshop erstellten Artefakte abgleitet wurden. Diese Artefakte spiegeln die Bedürfnisse und Probleme der Studierenden wider und dienten als Ausgangspunkt für die Entwicklung und Validierung der High-Fidelity-Prototypen. Der Prozess umfasste unter anderem die Erstellung und Validierung von Personas, Problem Statements, User Journeys und des erwähnten Prototyps.
Schlussfolgerung
Die Ergebnisse dieser Arbeit umfassen einen validierten Prototypen, der durch die neu eingebauten Funktionalitäten überzeugt. Darüber hinaus wird der Entwicklungsprozess reflektiert, und es werden konkrete Empfehlungen für die zukünftige Weiterentwicklung der Plattform formuliert. Mit dieser Arbeit wird ein menschzentrierter Ansatz für die Verbesserung der Plattform lost.university demonstriert, der eine Balance zwischen technischer Umsetzbarkeit und den Anforderungen der Studierenden schafft. Die Ergebnisse stellen eine solide Grundlage für die Weiterentwicklung und Implementierung dar
Photochef
No full textIntroduction: The Smarteating application is a web application that allows users to manage recipes in a smart way. Among its other features, the application is capable of converting photos of recipes to text and suggesting meal plans depending on the user's nutritional needs. The application, however, lacks essential user-centric features and an intuitive user experience.
The objective of the project is to survey and implement additional features that improve the usability of the Smart Eating application. These features include editing, saving, and sharing recipes, user authentication and user management as well as a new dashboard. Furthermore, the user interface (UI) should be improved to make the application more user-friendly and appealing.
Approach: To approach said features, widespread software practices are applied. Requirements engineering is conducted via user research. The problem domain is analyzed to
determine which parts of the application need modification and extension. After understanding how the application works and analysing the UI, mockups are created. Furthermore, additional features are discussed and planned for during the elaboration phase.
The smart eating web application offers a sophisticated and intricate structure. While documentation is limited, it presents an excellent opportunity for collaboration, encouraging productive discussions and meetings with advisors to ensure smooth and informed implementation.
Result: The possibility to log in/register through GitLab is added to the existing login window. The dashboard is also added along with mock data. User-Roles are added to the data structure and admin-users are now able to manage other users of the application in the newly added admin panel. Furthermore, it is now possible to edit the recipe
information that is extracted during the analysis of a recipe. The general UI appearance is redesigned through many changes including the shape of components, the colours, the general layout, etc. Finally, these changes are all verified by an intricate usability test to ensure that the changes truly improve the user experience and overall appearance
Text-to-SQL for DataGovernance Technologies and Education
No full textThis project explores the development of a Natural Language to SQL (NL-to-SQL) system using
Large Language Models (LLMs). The primary objective is to research different approaches of
generating SQL from NL and evaluate their feasibility. Subsequently a PoC was implemented to
demonstrate real-world usefulness.
The research found four key approaches which were evaluated: pure LLM, in-context learning,
fine-tuning the LLM, and Retrieval-Augmented Generation (RAG). Fine-tuning was ruled out due
to an insufficient amount of training data and time. The remaining three approaches were imple-
mented in Python and relevant LLM APIs and thoroughly evaluated.
Key findings indicate that while pure LLM approaches and in context learning provide a base-
line, RAG significantly enhances the accuracy and reliability of the generated queries. The results
of the test queries were evaluated in terms of 1. similarity (how similar is an output compared to
the example solution?), 2. validity (is the output valid SQL?), 3. executability (can the output be
executed and are the generated column names correct?), 4. reliability (how similar is the output
to the same user prompt?).
Testsets were divided into two grades: basic and advanced, based on its complexity. For the
advanced test cases, on average 45% can be executed on the databases, with only 70% of the re-
quested columns extracted. Hallucinations could not be completely eliminated when extending
the scope of the context, resulting in the low number of executable SQL queries. In this thesis,
the llama3.2 model was found to have the most potential for further development.
The PoC application demonstrates the feasibility of using RAG with metadata about database
schemas as well as JSONL input of users to generate SQL from NL, offering a user-friendly inter-
face for both technical and non-technical users
Arithmetik in endlichen Körpern
No full textDas Ziel der Arbeit war es, die Arithmetik in endlichen Körpern zu implementieren.
Verwendet werden sollte diese dann in einer existierenden Web-Anwendung zum Gauss-
Algorithmus (Gauss-Calculator). Die Implementation sollte dabei nicht nur den endlichen
Körper Fp unterstützen, sondern auch Erweiterungen davon, wie Polynomringe
Fp[x] und die Erweiterung um ein Minimalpolynom zu einem neuen Körper. Die existierende
Arithmetik der Anwendung verwendete den double-Datentyp, eine Approximation
des Körpers Q. Diesen Körper sollte man neu über die Benutzeroberfläche wählen und
konfigurieren können.
Um die Herausforderungen der Implementierung zu identifizieren, wurde zunächst eine
detaillierte Analyse der Problemdomäne durchgeführt. Anschliessend wurde die Implementierung
iterativ entwickelt, wobei ein Taschenrechner als erste Anwendung der
Arithmetik implementiert wurde. Die Funktionalität des Rechners wurde schrittweise
erweitert, um praktische Anwendungsfälle zu ermöglichen. Die Wahl der Reverse Polish
Notation (RPN) für den Rechner ermöglicht eine effiziente Umsetzung der mathematischen
Operationen. Die ursprünglich geplante Integration der entwickelten Lösung in
den Gauss-Calculator wurde dem Auftraggeber überlassen.
Es gelang eine geeignete Abstraktion zu finden, mit der die Arithmetik und Zahlenformate
der verschiedenen Körper und Ringe einfach mit der Benutzeroberfläche verbunden
werden konnte. Diese besteht hauptsächlich aus einer Grundklasse, die verfügbare
mathematische Operationen definiert. Implementiert werden diese für die einzelnen mathematischen
Körper und Ringe in ihren eigenen Klassen.
Das Resultat dieser Arbeit ermöglicht es mit langen Ausdrücken in verschiedenen
Körpern und Ringen zu rechnen und illustriert den Zusammenhang zwischen diesen
Exploring the use of Haskell to Program Microcontrollers used in Educational Robotics Platforms
No full textNowadays children are introduced to electronics and programming by developing applications for “educational robotic kits”.
These kits allow children to write simple applications for a robot that interacts with its environment. Because Python or a block based system like Scratch are beginner friendly, they are often supported by these kits. Sometimes event C/C++ SDKs are available for more advanced users. None of the common kits support a functional programming language and therefore the children do not have a chance to try a different approach to programming. Previous research has proved that it is possible to write a Haskell application on bare metal by implementing an operating system. With MicroHs, a newer Haskell compiler is developed, that is based on combinators.
The author of MicroHs has already demonstrated that it is possible to write simple MiroHs applications that run on micro controllers.
A different approach is used in the project Categorifier. It utilizes the "Compiling to categories" to transform Haskell code into C code. In this project, MicroHs is used to develop a line following algorithm that runs on a Raspberry Pi Pico.
The Raspberry Pi Pico controls a PicoGo robot to demonstrate, how a functional programming language can be used for an educational robotic kit. While it is possible to develop Haskell applications for devices with little memory, it is rather difficult as the Haskell runtime must be adjusted and a lot of C code is still required to get it on to the device
Integration of Deep Computer Vision Foundation Models for Document Interpretation and Anonymisation
No full textVisual Document Understanding (VDU) models, combined with Optical Character Recognition (OCR) or OCR-free, offer businesses and institutions a great opportunity to digitalise their processes and improve workflows. The digitalisation is progressing. However, challenges like sufficient knowhow to integrate VDU models, compliance with data protection regulations and identifying the processes, where VDU models offer the most significant benefit, have to be resolved.
The main goal of the work is to analyse and evaluate the practicality and appropriateness of available VDU models for processing of documents (e.g. PDF of scanned documents) and to demonstrate these in a Proof-of-Concept (POC) application. Even though some regulatory aspects, especially regarding anonymisation, are discussed in the work, the developed application does not aspire to be regulatory compliant.
During this work, two areas have been identified, where a tool to extract text from an image, identify relevant entities of personal information and anonymise these, is beneficial. First, the anonymisation of medical documents makes more data available for research and educational purposes. A second application is data leakage prevention, where detecting client data from screenshots would lower the risk of data breaches.
Various tools exist to extract text from an image. In the scope of this project, three tools have been integrated i.e., Tesseract, Amazon Textract and OpenAI GPT-4V(ison). The application extracts the text of uploaded documents or images and provides the user with the resulting text from all three tools. The user will be able to select the text with the best quality. Afterwards, a Named Entity Recognition (NER) Transformer model (i.e., bert-base-NER model) is used to identify the names of persons in the extracted text. The last step is the pseudonymisation of the entities. A randomly generated unique string replaces the entities in the text, so that a person cannot be identified based on the name in the text.
Another feature of the application is the evaluation of the OCR accuracy. The user is able to upload an additional ground truth file, which will then be compared with the output of the uploaded images. To calculate the OCR accuracy the Jaro Similarity string comparison algorithm is used. Furthermore, the NER model can also be tested by uploading the expected entities of the document in a separate file. The test will then show how many of the provided entities have been found in the extracted text.
It is impressive how powerful today's text extraction and NER models have become. However, during the work, it was recognised that they are not yet off-the-shelf and just ready to use. Neither works each tool perfectly, so errors are propagated to subsequent processes nor are the outputs of each tool standardised. To overcome such limitations, the process of text extraction and entity recognition should be executed by one model, which is also fine-tuned on the specific document types
Secure Device Provisioning Using SZTP
No full textInitial Situation
Managing many network devices takes a lot of effort and poses risks of inconsistency in configuration. Furthermore, time is needed to plug the device in, attach your computer and connect to the console for configuration. With automation, this process can be made much more efficient and reliable. The classical approach for this is called Zero Touch Provisioning (ZTP), meaning the device doesn’t have to be touched to configure it. Instead, it is registered in an inventory and can load a predefined configuration automatically.
Objective
The objective of this term project is to lay the groundworks for a network controller that handles zero touch provisioning of newly installed devices, as well as transferring configuration in case of 1-to-1 device replacement. The scope of the controller is limited to Cisco devices for this project, but it should be extendable to support various other devices. Furthermore, the controller should also be open for future extension, providing more functionality like ongoing configuration after the provisioning process.
Result
The authors developed a controller that bridges between inventory management and network device. On one end Netbox is used to manage device parameters and context-dependent configuration. Both are rendered into the target devices configuration using a template. On the other end the devices use DHCP to get SZTP redirect information for the controller. The provisioning controller provides an endpoint for SZTP-compliant devices to securely get the data needed to bootstrap themselves via HTTPS. The bootstrapping data includes firmware target version, download source and integrity hash as well as the configuration itself. During the devices lifetime our backup controller is used to automatically retrieve configuration backups. Nornir and Napalm are used to run the backup task on all devices registered in Netbox. In the event of a hardware failure, a replacement device can quickly be set up. By simply setting the configuration source device in Netbox the provisioning controller will automatically load the backup configuration
ChatGPT-3 Security Labs
No full textAI is currently a topic that is being addressed in most industries. This is no exception, especially for cybersecurity. Advantages could be realized, for example, in the automation of security analysis and thus in the relief of analysts. As is often the case, there are disadvantages here too. The use of AI makes it possible for a wide audience to achieve rapid success without in-depth knowledge.
The aim of the work is to research the effects of AI on cybersecurity. For this purpose, both the Blue Team side and the Red Team side are explored. Ultimately, the aim is to create labs that students can use to work on.
The topics were selected using a decision matrix with weights based on various criteria. This is intended to address the most relevant topics. A Proof of Concept (PoC) was then carried out for each selected topic. Topics have developed in the categories of reverse engineering, exploiting, threat modelling, phishing and prompt engineering.
In our research, we focused on prompt engineering, a recognized area in AI where control over the AI's output is consciously managed. Our experimentation revealed that retraining the AI through prompt engineering was facilitated more effectively using the OpenAI API, which allows the provision of "system content" to influence the AI's behaviour. Utilizing DAN as system content in our Hacking Lab, we successfully conducted various experiments, making potentially malicious requests to the AI without encountering blocking mechanisms. Our results, illustrated in the evaluated labs, showcase the efficacy of prompt engineering in achieving desired outcomes
Konzeption und Umsetzung eines ML-basierten Recommender-Systems für Freizeitaktivitäten
No full textEinleitung: Diese Arbeit beschäftigt sich mit der Entwicklung eines personalisierten Empfehlungssystems für Freizeitaktivitäten, das psychologische Merkmale der Nutzer (Big-FiveModell) mit standortbasierten OSM-Daten kombiniert. Anders als herkömmliche Verfahren, die umfangreiche Nutzerinteraktionsdaten erfordern, stützt sich unser Ansatz auf wissenschaftlich validierte Persönlichkeitsprofile und ein kontinuierliches Like-/Dislike-Feedback zur Feinjustierung der Empfehlungen.
Vorgehen: Zunächst wird ein orthogonal gestalteter Fragebogen eingesetzt, um die anfänglichen Nutzermerkmale zu erheben und daraus einen Nutzervektor abzuleiten. Das Empfehlungssystem nutzt einen zweistufigen Algorithmus: Phase 1 (Klassenbasiert) für Kaltstart und wenige Tag-Learnings, Phase 2 (OSM-Tagbasiert) für eine feingranulare Anpassung, sobald genügend Tag-Vektoren vorliegen. Hierbei misst ein Matching-Algorithmus die euklidische Distanz zwischen Nutzer- und Aktivitätsvektoren. Die technische Umsetzung umfasst ein Flask-Backend (Datenverarbeitung und Kommunikation mit der Overpass API) und ein React Frontend (UI für Suche, Feedback, Visualisierung). Eingebundene Echtzeit-POI-Daten aus OSM werden dynamisch klassifiziert und in unsere Empfehlungslogik integriert.
Ergebnis: Die Evaluierung verdeutlicht, dass das System bereits mit begrenztem Feedback relevante und stimmige Empfehlungen erzeugt. Der zweistufige Ansatz ermöglicht robuste Ergebnisse bei Kaltstarts (Phase 1), während er bei weiterem Feedback (Phase 2) immer individuellere Vorschläge liefert. Die kontinuierliche Anpassung der Nutzerprofile über Like-/Dislike-Eingaben steigert die Personalisierung schrittweise. Erste Tests bestätigen die Lernfähigkeit und Stabilität des Systems. Zukünftige Weiterentwicklungen könnten kontextbasierte Faktoren (Wetter, Events) oder ein federiertes Lernverfahren integrieren, um Privatsphäre zu erhöhen und das Empfehlungsspektrum weiter zu verfeinern