Security and trust
How GPU Goblin protects your machine and checks results, and what is not protected.
On your computer
- No foreign code. The software only runs GPU Goblin’s own computation code. A task is a description (a model’s layers, a batch of data, a size).
- A sandbox for every task. On Linux, each task runs isolated, with no access to the network or to your files, with memory and disk limits. The software refuses to start if one of these protections is missing.
- Checked data. Before computing, the software checks the server’s signature when the data arrives. If a piece of data has been modified, the computation is rejected.
- Local control. The software can only be controlled from your own machine.
On the results
- Computation checks. Each result is either recomputed by an independent validator or computed twice by two different contributors and compared within a measured tolerance.
- Spot checks. A share of the results is recomputed a third time, at random, even when two contributors agree.
What is not protected
- Contributors receive pieces of the data they compute: arrays of numbers without names or descriptions, only what each task needs. The data is not encrypted during the computation: do not send confidential data to the open grid.
- Your public address is visible to the other machines authorized on the same data when sharing between contributors is on (it is, by default, and can be turned off).
Reporting a problem
Found a security issue? Write to contact@gpugoblin.com before publishing it.