Bachelor and Master Theses

Background:

Smart-grid and IoT infrastructures rely on distributed, networked embedded devices that must process data locally while ensuring privacy and reliability. Federated Learning (FL) allows these devices to collaboratively train a shared machine-learning model without transmitting raw data to a central server.
While privacy is preserved, FL introduces communication overhead—each client repeatedly exchanges model updates with a server. For embedded and edge systems, this additional traffic can affect latency, scalability, and energy use. Understanding these trade-offs is critical when deploying FL in cybersecurity domains such as network-based intrusion detection.

Aim:

To implement a simple federated-learning setup for intrusion detection and evaluate its communication efficiency, network load, and convergence performance across different configurations.

Main Tasks:

• Implement a federated-learning environment using Flower or TensorFlow Federated (TFF).
• Use a public intrusion-detection dataset to train a small model (e.g., MLP or CNN) within the federated setup.
• Measure key performance indicators:
  • Number of communication rounds to reach target accuracy
  • Total bytes exchanged between clients and server
  • Latency and runtime per round
  • (Optional) CPU or memory usage per client
• Analyze results (e.g., accuracy vs rounds, data vs performance).

Possible Extensions (Optional for a team of two student):

• Compare FedAvg and FedProx aggregation algorithms.
• Evaluate IID vs Non-IID data distributions.
• Study the effect of model size on communication cost.
• Create a small visualization dashboard (e.g., Streamlit) for real-time monitoring.

• Skills: basic Python programming; interest in distributed systems and network performance
• Familiarity with ML frameworks (e.g., TensorFlow)
• No deep-learning expertise required—existing frameworks will be used.