Chiron: Privacy-preserving machine learning as a service

Researchers designed and implemented Chiron, a machine learning as a service system that protects training data privacy and model confidentiality under the premise that the cloud service provider is untrusted. Chiron places the standard ML training toolchain (including the Theano framework and C compiler) inside an SGX enclave and uses the Ryoan sandbox to further restrict the service provider's model creation code, preventing training data from leaking outside the enclave. To support distributed training, Chiron runs multiple concurrent enclaves that exchange model parameters through a parameter server. Researchers evaluated it on image classification benchmarks such as CIFAR and ImageNet, claiming that its training performance and the resulting model accuracy are practical for common machine learning as a service use cases.