Theses

(Student thesis, Oct. 2025 - Mar. 2026, Student: Xi Cheng) 

Denial of Service (DoS) and Distributed Denial of Service (DDoS) attacks pose a critical threat to network infrastructure by overwhelming targets with excessive traffic, resulting in severe service disruptions and significant reputational damage for organizations. Machine learning (ML) techniques have shown great potential for detecting such attacks; however, traditional ML-based solutions typically rely on centralized data collection and processing, which can introduce significant network overhead and even create bottlenecks under large-scale attack traffic. Recent advancements in programmable data plane technologies have created new opportunities to deploy ML-driven detection and mitigation mechanisms directly within the network. This project focuses on designing and evaluating lightweight ML-based mechanisms for detecting and mitigating DoS/DDoS attacks using software-based programmable network emulators. The student will implement and optimize ML models for deployment on P4-based data planes using BMv2 and emulate realistic network topologies in Mininet.

(Diploma thesis, Aug. 2025 - Jan. 2026, Student: Yijun Zhu) 

Control algorithms for robot swarms must be tested under realistic wireless conditions before they are deployed in the field. Robots’ motion and the wireless channel are mutually dependent: every action changes distance, shadowing and multipath; link quality then determines which observations arrive on time; those observations steer the robots’ next motion. Control and perception therefore rely on real-time bandwidth and latency. Single-domain simulators ignore this feedback: physics engines such as Gazebo deliver photorealistic sensors but assume perfect links; ROS 2 provides convenient message passing yet hides true network behaviour; network simulators like ns-3 model are blind to kinematics. Without a repeatable, closed-loop environment, latency- or loss-induced failures remain unnoticed until costly field trials. To bridge this gap, the student will first build a closed‑loop co‑simulation platform. 

(Diploma thesis, Jan. - Jul. 2025 Student: Chenjun Wu)

Machine learning has become an essential tool for solving networking challenges, such as traffic classification, anomaly detection, and network configuration. By leveraging machine learning algorithms, networks can dynamically manage data traffic and optimize overall performance. Recently, in-network machine learning solutions have emerged, driven by the evolution of network devices that are both high-performance and programmable. Technologies like Switch ASICs, network interface cards (NICs), and FPGA-based network devices now utilize P4, a domain-specific language that enables the definition and customization of network protocols and processing functions directly within the data plane. This level of programmability opens new possibilities for offloading traffic classification tasks to network hardware. In this project, students will identify an appropriate machine learning algorithm suited for network communication scenarios and deploy it on a Switch-ASIC (Tofino Switch) testbed. The goal is to enable real-time detection and classification of network traffic conditions directly within the switch hardware.

(Diploma thesis, Jun. - Nov. 2024 Student: Yixin Yuan) 

With the development of science and technology and the application of AR, VR, and other technologies, multimedia data streams, such as video, audio, and haptic, are increasingly prevalent in networks. These applications have higher requirements for latency and synchronization. To address the latency problem, network slicing has been proposed. However, with differentiated services, the multimedia streams can be out of sync. Furthermore, the expansion of network capacities is accompanied by increased buffer sizes at each network node. Excessive buffer occupancy can lead to buffer bloat, a condition where the increased latency and reduced throughput undermine network performance. To tackle these multifaceted challenges, a novel scheme will be proposed for addressing the critical synchronization issues present in modern network operations.

(Diploma thesis, Mar. - Sept. 2023, Student: Jiakang Weng) Each service in a cloud-native application communicates with one another via a software proxy called sidecar. A sidecar intercepts cloud traffic reaching service and thus provides various control functions such as security and traffic management. However, as each sidecar is co-located with each service, this design introduces overhead (e.g., increasing latency and lower throughput) for cloud-native applications, especially for applications that include many services. This work aims to improve the performance of service mesh for cloud-native applications.

(Diploma thesis, Jun. - Nov. 2023, Student: Junyue Wang) 

Applications are classified into stateless and stateful. Stateless applications do not require acknowledging application states (i.e., historical processing values) to handle users’ requests. Meanwhile, stateful applications rely on application states for proper processing. More importantly, application states are used to provide the fault tolerance and scalability of applications that require state transfer between applications.

Existing studies have been proposed to utilize the states of applications deployed on general-purpose servers. Due to the emerge of latency-sensitive use cases such as autonomous driving and robots, there is a possibility to deploy applications directly on programmable network devices such as Tofino switches or SmartNICs, thus allowing applications to process users‘ requests at line rate and consequently reducing latency. While ensuring line-rate processing for applications, programmable network devices have to maintain application states, thus making fault tolerance and scalability challenging. There is also another possibility that uses programmable network devices to accelerate state transfer. This work aims to investigate a solution for the state transfer of stateful applications using programmable network devices

(Master thesis, Jun. - Nov. 2023, Student: Um e Habiba)

Unlike traditional monolithic applications, cloud-native applications are the collection of small and independent services, which are so-called microservices. As cloud-native applications have gained tremendous interest in recent years, many cloud vendors such as Google Cloud and Amazon Web Service already provided cloud platforms for cloud-native applications. Cloud-native applications are realized as containers, which are provisioned by container orchestration platforms, such as Kubernetes. Telecom operators such as AT&T advocates containerizing their existing applications. However, this containerization processing comes with challenges. First, for the security concern, many existing applications run on top of virtual machines. KubeVirt, which is the extension of Kubernetes, has been introduced to orchestrate both containers and virtual machines. Second, to better utilize hardware capabilities, many existing applications run on bare-metal servers. Metal3, which is also the extension of Kubernetes, has been proposed to orchestrate both containers and bare-metal hosts. However, the bare-metal hosts are typically equipped with CPUs that introduce lower processing rate than emerging hardware acceleration technologies such as DPDK and SmartNICs. This topic investigates the performance of cloud orchestration under the support of hardware acceleration technologies.

(Research Project NES-12 PW-14.1, August 2023, Student: Josef Gudnason) This project work aims to develop and evaluate communication protocols between AR glasses and FingerTac (an embedded system providing the feel of touch), considering delay-sensitive applications while efficiently using network resources.

(Diploma thesis. August 2023. Student: Mingyu Ma) Network infrastructures usually must deliver data flows from many applications with different QoS requirements. Some applications are bandwidth-intensive, while others are latency-sensitive. Therefore, network nodes must allocate (switching) resources wisely to meet different requirements. Software-Defined Networking defines and applies rules for various data flows. However, these solutions suffer from high memory and overhead as the number of flows increases. Algorithms at the data plane can resolve this issue when applying programmable switches. For example, switches can apply priority rules on packets depending on their expected deadlines. State-of-the-art (SoTA) tend to drop packets that fail to meet their deadlines to slow the source’s sending rate. However, this strategy fails to support low packet rates and delay-sensitive applications like haptic teleoperation. However, the stringent latency and low loss requirements demand reexamining the state-of-the-art and developing novel algorithms. This diploma thesis aims to develop a novel algorithm prioritizing delay-sensitive applications while maximizing network resource utilization.

(Student thesis, July 2023, Student: Yixin Yuan) In this study, we focus on collecting and analyzing practical network traffic data generated from users performing a pre-defined set of teleoperation tasks using Haptic stylus pens in a simple, dedicated Local Area Network without interference. The goal is to understand the characteristics of the data stream at the output of the perception-based haptic codecs.

(Research Project INF-PM-FPA, March 2023, Student Fritz Windisch, co-supervised with Prof. Strufe KIT) Resilience and reliability of network infrastructure play a more important role in our lifes every day. Due to this, evaluation of resilience and reliability of network infrastructure should be a priority in development. To test network infrastructure against attacks in a partially virtual environment, we create a network testbed to simulate SDN networks featuring decentralization and inte- gration of real-world hardware.

(Student thesis, July 2022, Student: Tobias Scheinert) This student thesis aims to develop a novel algorithm prioritizing delay-sensitive applications while maximizing network resource utilization.

(Student thesis, March 2023, Student: Kunru Zou) This student thesis investigates the 5G-TSN integration more thoroughly, including getting familiar with TSN, 5GS, and OMNeT++ via literature study, setting up a testbed based on OMNeT++ network simulator, developing 5G-TSN integration, and finally evaluating the system's performance.

(Student thesis, August 2023, Student: SophianRomdhani) This study involves reviewing state-of-the-art and key challenges on D2D communication protocols in the context of 5G, investigating the potential benefits of D2D communication in various applications, measuring Key Performance Indicators (KPIs), and analyzing network performance when employing D2D communication utilizing the OMNeT++ simulation software.

(Research Project CMS, Jan. 2024 - April. 2024, Student: Bharat Veauli)

With Multiplayer virtual reality(VR) game growth, understanding the intricacies of encrypted game traffic has become imperative, not just from a network optimization standpoint, but also from privacy and security perspectives. This project delves deep into this domain, aiming to unlock patterns and insights from encrypted data streams of popular multiplayer games- from identifying specific game types being played to even pinpointing user activities within those games.

(Research Project MT-12-STA, Dec. 2023 - May. 2024, Student: Guangchen Zhu)

The evolution of Augmented Reality (AR) and Virtual Reality (VR) has ushered in a new era of immersive experiences, a critical component of this immersion is the haptic feedback, which provides tactile sensations to users, making virtual interactions feel "real". FingerTac is a wearable device for a haptic thimble. It applies simultaneous vibrations to the finger and produces tactile feedback at the bottom center of the fingertip. Controlling the frequency of the FingerTac's vibration efficiently, and eliminating redundant signals can improve the overall user experience and reduce computational overhead and latency.

This project work aims to redesign FingerTac's codebase for develop three distinct vibration patterns (from soft to strong), and optimize signal transmission to reduce packets and latency. Finally, the comprehensive communication evaluation will be conducted.

(Diploma thesis, Nov. 2023 - May. 2024, Student: Yuzhe Wang)

In the field of environmental monitoring and industrial safety, the rapid and accurate classification of gas mixtures is important for ensuring air quality and detecting hazardous substances. Machine learning has become a widely adopted method for such classification tasks, offering the ability to discern complex patterns in large datasets. The metal oxide gas sensors, deployed across various locations, generate substantial amounts of data that require transmission to a central system for analysis. However, the volume of this data brings significant challenges: its centralization of transmission and processing can be inefficient and time-consuming. To address these issues, a distributed computing can be a potential approach, facilitating the processing of data across multiple nodes and thereby accelerating the analysis process.

To successfully complete the thesis on this topic, the student must simulate a distributed approach. This involves providing a comprehensive understanding and evaluating the performance in the scenario of mixed gas separation.

(Master thesis, April-September 2023, Student Fritz Windisch, co-supervised with Prof. Strufe KIT) Tele-operations require secure end-to-end Network Slicing leveraging Software-Defined Networking to meet the diverse requirements of multi-modal data streams. This requires a flexible yet robust and resilient provisioning of slices spanning multiple autonomous networking domains. Furthermore, studies in this research area need tools to develop prototypes quickly that work on emulation and practical deployment. This thesis aims at developing resilient micro-slices. The proposed solutions will be examined on a realistic testbed to compare with state-of-the-art.

(Diploma thesis, Jun. - Nov. 2023, Student: Yankai Su) 

Network gateways usually must deliver data flows from many applications with different QoS requirements. Some applications are bandwidth-intensive such as videos, while others, such as audio and haptic, are latency-sensitive. Therefore, network gateways must allocate (switching) resources wisely to meet different requirements. Software-Defined Networking defines and applies rules for various data flows. 

This diploma thesis aims to develop a novel algorithm that leverages the differences in sensitivities of human brains to particular modalities. The goal is to maintain low latency for the haptic data stream while minimizing the quality degradation of other modalities and maximizing network resource utilization.

(Diploma thesis, Nov. 2023 - Apr. 2024, Student: Yifan Ma)

Human-machine interactions can benefit from human gesture recognition. Conventional methods, such as visual-based and wearable devices, usually suffer from obstructions and inconvenient wearables. Radar-based recognition has clear advantages in hands and fingers free and natural motion. This approach offers convenience and efficiency, particularly in complex environments. State-of-the-art in the area has employed mainly neural networks, such as the convolutional neural network (CNN). Some combined CNN with LSTM networks by slicing spectrograms into time series. Alternative machine learning techniques such as k-NN have also been explored. However, neural networks have dominated the gesture recognition methods due to their remarkable performance. The diversity of methods and ununified input data hinders comparing the existing method's effectiveness and accuracy. Therefore, the primary objective of this thesis is to develop a benchmark platform to compare recognition schemes.

(Diploma thesis, Jul. 2024 - Dec. 2024, Student: Yuli Jiao)

Traditional methods in human gesture recognition, like visual-based and wearable devices, often face obstacles, are inconvenient to wear, and raise privacy concerns. Radar-based recognition has clear advantages as it allows for hands and fingers to be free and for natural motion. This approach offers convenience and efficiency, especially in complex environments. A previous pipeline was developed forto assess the effectiveness of various deep learning models and datasets. However, the previous work has suffered from degraded accuracy and limited datasets. 

This thesis aims to improve that pipeline in various aspects, like automated hyperparameter tuning (HPT), various data transformation methods, additional datasets, and deep learning models. One of the main contributions is the HPT function, which enhances model performance by optimizing hyperparameters. Comparisons indicated that the modified pipeline significantly improves recognition accuracy. These enhancements enable a more comprehensive evaluation of radar-based hand gesture recognition.