Description
The adoption of Internet of Things (IoT) devices in homes and other buildings is increasing in popularity.
However, many people are afraid they could be surveilled by their smart home applications, as these
generate privacy-relevant data that attackers could access by targeting them. Protecting such data remains
a significant challenge. One possible solution is to use an Intrusion Detection System (IDS) that learns
normal and malicious behaviour from a dataset to detect attacks. However, a key challenge is the lack of
realistic, privacy-aware (i.e., containing privacy-sensitive data) datasets suitable for training these models.
This thesis addresses this gap by constructing a smart home testbed based on three Raspberry Pi 5
devices running the smart home management platform Home Assistant, generating normal network traffic
and simulating seven attack types. A privacy analysis reveals behavioural inference risks in the raw data.
To mitigate these, the privacy-enhancing technology (PET) synthetic data is used, producing a statistically
equivalent, privacy-preserving dataset used for Machine Learning (ML) training. Afterwards, three ML
models, Random Forest, XGBoost and Multilayer Perceptron, are evaluated using the Train Synthetic, Test
Real (TSTR) methodology. The results show a detection performance loss of less than 1% compared to
real data, indicating that synthetic data is a viable PET for privacy-preserving IDS training in smart building
environments.
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching