Detektering av utom-distribution inom federerad inlärningsmiljö via hyperdimensionell funktionsfusion

Federated Learning (FL) offers a promising framework for privacy-preserving machine learning by enabling decentralized model training across multiple clients. However, it remains vulnerable to data heterogeneity and malicious participation, particularly in the form of Out-of-Distribution (OOD) data and poisoning attacks. This thesis proposes a lightweight and privacy-preserving OOD detection mechanism based on Hyperdimensional Feature Fusion (HDFF), leveraging principles from Hyperdimensional Computing (HDC). By projecting neural network feature vectors into a shared high-dimensional space using semi-orthogonal matrices, model representations are transformed into hypervectors that can be reliably compared using cosine similarity. This approach enables the detection of anomalous or adversarial client models without accessing raw input data or internal states, maintaining full compliance with FL privacy constraints. The system is implemented in a custom-built local FL simulation, and evaluated across various OOD scenarios, including label-fipping and unseen data modalities. Experimental results demonstrate that the proposed method achieves high accuracy and AUROC in distinguishing OOD models while preserving global model performance. The findings suggest that HDC-based representation comparison is a viable and efficient solution for enhancing trustworthiness in FL systems.