MIND: A Privacy-Preserving Model Inference Framework via End-Cloud Collaboration

Image processing and analysis capabilities have been revolutionized by the rapid advancement of deep learning technologies. Model inference enables powerful image analytics, however, complex neural networks often require more computing and storage resources than end device can provide. While cloud-based model inference helps alleviate the resource constraints of these devices. However, transferring sensitive image data to the cloud server raises significant privacy concerns. Existing approaches usually either encrypt the images locally and send them to the cloud server or deploy the model locally, which fails to exploit the computing capability of the end device or reveals a machine-learning model. To this end, in this paper, we proposed MIND, a privacy-preserving model inference framework via end-cloud collaboration. MIND makes the end device and the cloud server jointly perform layer-by-layer computations required by model inference through cryptographic primitives including the homomorphic cryptosystem and additive secret sharing. For linear layers, MIND always chooses the less expensive option between data encryption and model parameter encryption. Results of experimental evaluations on several models show MIND is effective and outperforms the existing solutions.