Federated Learning in Edge Devices: Strengthening Data Privacy and Security
Master of Applied IT
Project description
This study investigates performance-based dynamic clustering in Federated Machine Learning (FML) for possible usage in autonomous vehicles, focusing on mitigating data poisoning in resource-constrained edge environments. By grouping clients based on performance metrics, the approach enhances model robustness under varying adversarial conditions, including poisoned clients and diverse data distributions. Using MobileNetv3, the results demonstrate improved accuracy and resilience, offering insights into securing FML systems.
Context
FML enables decentralized training, addressing privacy concerns and reducing cloud dependency. To tackle challenges like resource constraints, Non-IID data, and security threats such as data poisoning, the study employs dynamic clustering based on client performance. Using MobileNetv3 for image classification, simulations demonstrate that clustering enhances model accuracy and resilience under adversarial conditions, such as poisoned data and diverse distributions. While effective, limitations include clustering instability in extreme cases, highlighting areas for refinement. This research advances scalable, secure, and privacy-preserving machine learning for autonomous systems.
Results
This study explored the application of dynamic clustering as a mitigation strategy against data poisoning in FML (FML) systems deployed on edge devices, with a focus on autonomous vehicles. The tests conducted revealed that clustering based on performance metrics can reduce the impact of poisoned updates in multiple scenarios by 9.05% on average or 13.74% in best case scenario. The use of
MobileNetv3 as the selected model demonstrated the feasibility of deploying lightweight and resource-efficient machine learning solutions for real-time environments.
Computational analysis confirmed that the proposed approach operates well within the capabilities of modern automotive-grade CPUs and GPUs, reinforcing its practicality for real-world applications. These results underscore the potential of dynamic clustering to enhance the security and robustness of FML systems.