Contribution Scores in Private and Robust Federated Learning
Federated Learning is a very young discipline, and its security and privacy problems are under heavy research scrutiny. Contribution scores are also a fundamental interest for the scientific community and the R&D companies involved with this cutting-edge technology. Consequently, the field where these two directions overlap is primarily unexplored and contains crucial research questions. This is not surprising, as the goals of these directions are seemingly contradictory.
- The main goal of privacy in federated learning is to hide exact details about the individual datasets without severe ramifications concerning the learning objective.
- The main security goal in federated learning is to control the individual influence on the shared model to prevent any malicious activity originating from an adversary.
- The main goal of contribution scores in federated learning is to uncover fundamental details about the participants’ datasets to sufficiently differentiate them.