Federated Learning - Christophe Cerisara

HeteroFL: central model but with heterogeneous devices/capabilities/topologies
Assisted Learning: no central model; similar to stacking but does not transmit the labels, only the local predicted errors. I see 2 main drawbacks:
- it assumes that every X has a unique ID, and occurs in several nodes: this is fine for symbolic items that are few (e.g., citizens), but this does not hold for observations sampled from continuous density, such as speech signals, or combinatorial observations, such as goban positions or NLP sentences.
- it assumes that transferring the ID publicly is better than transferring the model: this may be wrong: knowing that a given citizen occurs in 2 health-related DB is a privacy threat. Also, knowing that the ID belongs to a training dataset is the outcome of membership inference attacks, and it's given straight away here. It may be better to rather make the model public, as it may be somehow protected against such attacks. In other words, this method is good when the knowledge of X is a threat, while when the knowledge of X is harmless, but knowing that X is part of the training set is a threat, then it may be better to make X and the models public.
Blockchain-FL: no central model
Google's FL comic