Why don't people speak of the vulnerable side of Federated Learning here
Y'all see that it protects the privacy of the client but does it actually? Attackers can try to hack the model and reverse engineering might be possible to rebuild the data. Not just that, attackers even manipulate the model by either poisoning the data or the model. Some use multiple accounts to shape the model their own way.
Their are many ways the models can be hurt to avoid convergence, so how do they actually protect their models? I would love to discuss it with people who have read about this.
Thank you