Encrypted machine learning, made simple
Bring your own PyTorch model. eML turns it into an encrypted model and runs it on GPUs. Your data is never decrypted, not even while it is being processed.
From PyTorch to encrypted in three steps
Bring your model
Start from the PyTorch model you already have. No cryptography expertise needed.
eML compiles it
Our compiler toolchain, built around Google's HEIR, translates your model to the encrypted domain.
Run it encrypted
Inference runs on our GPU library. Your inputs and results stay encrypted the whole time.
Fast enough to use. Simple enough to try.
Fastest encrypted inference
ResNet-20 runs fully encrypted in 139 ms on our GPU library, with no loss in accuracy.
No setup
A managed backend on NVIDIA GPUs. No CUDA install, no servers to run.
Private by design
Data is processed with fully homomorphic encryption. The server does the work but never sees your data.
Try it yourself
Run your model under encryption
eML is in beta and works best today with small convolutional networks. Support for larger models is on the way. Does your model not fit yet? Talk to us and we will help.