SecureQNN: Introducing a Privacy-Preserving Framework for QNNs at the Deep Edge

Recent concerns about real-time inference and data privacy are making Machine Learning (ML) shift to the edge. However, training efficient ML models require large-scale datasets not available for typical ML clients. Consequently, the training is usually delegated to specific Service Providers (SP), which are now worried to deploy proprietary ML models on untrusted edge devices. A natural solution to increase the privacy and integrity of ML models comes from Trusted Execution Environments (TEEs), which provide hardware-based security. However, their integration with heavy ML computation remains a challenge. This perspective paper explores the feasibility of leveraging a state-of-the-art TEE technology widely available in modern MCUs (TrustZone-M) to protect the privacy of Quantized Neural Networks (QNNs). We propose a novel framework that traverses the model layer-by-layer and evaluates the number of epochs an attacker requires to build a model with the same accuracy as the target with the information disclosed. The set of layers whose information makes the attacker spend less training effort than the owner training from scratch is protected in an isolated environment, i.e., the secure-world. Our framework will be evaluated in terms of latency and memory footprint for two ANNs built for the CIFAR-10 and Visual Wake Words (VWW) datasets. In this perspective paper, we establish a baseline reference for the results.