Interesting and challenging article. Food for thought: it shows the exit doors but not yet how to find the keys to open them. I see for example:

• homomorphic encryption to solve privacy: this encryption technique allows only some arithmetic operations on the items, not all operations.
• federated learning to solve influence: distributing learning will use HUGE amount of network traffic.

Let’s keep working… decentralised deep learning (not really AI :-) is not yet for tomorrow.