mofanv / darknetz

runs several layers of a deep learning model in TrustZone
MIT License
85 stars 29 forks source link

Meaning of protecting layer #25

Open aghia98 opened 3 years ago

aghia98 commented 3 years ago

Hi, I am studying DarkneTZ for my end-of-studies project.

I am wondering what are you exactly protecting when you put a layer inside an enclave: does its parameters becomes unreadable for external users? what about its gradients?

If gradients of protected layer are secured, what computation of backpropagation are you protecting?

Thanks in advance

mofanv commented 3 years ago

Hi @aghia98 , DarkneTZ is designed to defend against membership inference attacks, so the last several layers are in TEE, and both parameters and gradients cannot be accessed from outsiders. You can have a look at the corresponding paper to understand the backyard https://arxiv.org/abs/2004.05703

For fully protecting all layers, you can further check https://arxiv.org/abs/2104.14380