Towards Model-Centric Security for IoT Systems

In this paper, we make a case for a novel model-centric security approach to the IoT application systems. We thus depart from the popular device-centric and data-centric schemes. Our proposal is based on the premise that the trained model, rather than the fine-grained input and output data streams, plays the pivotal role in many IoT application systems. We thus seek to obfuscate the model directly, but not the individual data items or sensory data streams. We present our initial design of sampling-based model obfuscation. Both evaluations and analysis have partially confirmed our design to date.