Safeness

Training safeness

When trainers connect to a host training to report a model performance (In case of genetic algorithms) there could be the case that the trainer tries to fake the execution and reports something misleading about the real performance of a model.

For this reason EzSpark adopted a strategy to assure the safeness of the training. the safeness parameter that can be set in the configuration for the host training lets save a history host-side for each genome action during the training.

Higher is the safeness parameter bigger will be the history. Once a training ends a genome performance report the hosts give the task to other trainers to validate the history. In this sense the communication can be seen as a sort of blockchain where miners are the trainers that validate the host history

Host safeness

Thanks to the tunneling approach we adopted, people cannot perform a ddos attack to hosts. The only issue a host can encounter is to fullfill its local machine resources, but for this reason a max_number_of_trainers parameter has been added to python host code