Abstract

As the number of smart environments is increasing, our reliance on IoT devices and sensors is also increasing. However, the data from sensors may not always be reliable as sensors can report incorrect sensor readings and event notifications due to sensor failure or a compromise by a malicious actor. Although there has been extensive research on sensor data verification, they have their own limitations. Most of them deal with only certain types of sensor data, either only verifying the events notifications or only the sensor readings. They use redundant sensors for verification which might incur additional cost and overhead. As those works are designed to learn from a single smart home data without collaborating with other smart homes, they cannot utilize more diverse data from multiple smart homes to achieve better accuracy.
In this thesis, we present an approach that learns the relationships among the co-located sensors and uses that relationship to verify the reported sensor readings, and detects event notifications (including masked and spoofed events) by monitoring co-located IoT devices and also learns collaboratively from the data of multiple smart homes without sharing the private data to a centralized server using federated learning. We implement our solution in the context of smart homes and evaluate its effectiveness using a public smart home dataset. For sensor reading verification, we achieve an R2 score of 0.98, and for event verification, we achieve an accuracy of up to 100% which is among the best of existing works.