ImageBox: Small-Scale Blockchain-Aided Trusted Machine Learning Inference

Over the recent years, machine learning (ML) has gained more and more importance in a growing amount of areas. One of those scenarios is using ML in small-scale deployments on the edge, potentially even in remote areas without a stable internet connection, in which sensor data are to be processed. However, there are issues concerning trust that remain open and pose danger to those relying on the fast growing technology.This paper presents ImageBox, a framework for trusted ML inference, protecting the inference process from manipulation by malicious attackers. This is achieved by leveraging a Trusted Execution Environment (TEE) in combination with tamper-proof storage to store models, inputs and outputs. We present an implementation that leverages a blockchain framework that is executable on one singular node, without compromising security. Using this implementation, we show that ImageBox can make ML inference trusted with a performance overhead of just 35% compared to untrusted execution.