[gormanc]

According to the website (https://vectordash.com/hosting/) they use a highly isolated Ubuntu image, so the person hosting the service shouldn't have access to the VM with your model or data on it. It would be nice if there was some third party audit of the software though, the models, the code, and even the training data can be pretty sensitive for researchers.

[Samin100]

If your training data is sensitive, then Vectordash may not be the best GPU provider. But if you're a broke CS student like me who wants to participate in a few Kaggle competitions (after having burned up their AWS student credits in 3 days) without shelling out a bunch for a K80, then Vectordash might be pretty helpful!

[throwaway2048]

there is no way to "highly isolate" a VM from a host.

[lostmsu]

But there is (though I think they don't use it): TPM based host attestation.

[throwaway2048]

The microsoft secureboot golden key got leaked, anything based on secureboot as a root of trust is 100% blown wide open.

https://web.archive.org/web/20170604013028/https://rol.im/se...

[lostmsu]

I am not sure this depends on TPM. Care to share a link?

[JensRex]

If you don't want to claw your eyes out while reading:

https://bpaste.net/show/571ef50296ac

[q3k]

Theoretically possible via SGX.

[mmozeiko]

Which can be defeated with SgxSpectre: https://arxiv.org/abs/1802.09085

[snuxoll]

Oh goodie, I wonder if Netflix is going to disable 4K support on PC as a result of this (the requirement for Skylake was due to SGX).

[gruez]

Worthless if the GPU doesn't have something similar. Otherwise you can monitor the pci-e lanes for all the data the cpu is sending over to the gpu.