Implementing end-to-end test with Byzantine-Robust Aggregator for cifar10 task

[francesco98]

Discojs

• Changed cifar10 task definition to comply the input size required by the network. Indeed, MobileNet is using an input size of 224x224x3, while the example data (and the previous specified input shape) was 32x32x32.
• Preprocessing function to normalize tabular data has been added

Server

Changed the end-to-end test case for the federated setting to use both standard aggregator and byzantine-robust one. Also, the test is now considering the presence of multiple users, splitting the training data among them and validating the model computing the accuracy after the training phase. As last step, an assertion has been added to expect a reasonable accuracy after training.

[martinjaggi]

added some comments. one goal to keep in mind is to make the Byzantine robust clipping such that it can later also be used in decentralized. so let's try to define the clipping center based on local info only (and try to minimize additional communication if any).

we should also write down some thoughts&docu on how it will work with async (whenever the buffer size is reached, we do an aggregation with that group, so that's a round)