PoCoFL: POlicy-COmpliant Federated Learning
PoCoFL adds verifiable policy checks to federated learning without tying them to one topology or protocol.
The paper frames the gap as two-sided: clients may submit non-compliant updates, and aggregators may mishandle admitted ones. PoCoFL separates FL type, policy semantics, and cryptographic realization, then uses commitments and non-interactive zero-knowledge proofs for compliance. The authors instantiate it for vanilla, continual, personalised, and threshold-encrypted federated learning, and report proof-of-concept implementations for all four. ArXiv · AI/CL/LG's note
The paper frames the gap as two-sided: clients may submit non-compliant updates, and aggregators may mishandle admitted ones. PoCoFL separates FL type, policy semantics, and cryptographic realization, then uses commitments and non-interactive zero-knowledge proofs for compliance. The authors instantiate it for vanilla, continual, personalised, and threshold-encrypted federated learning, and report proof-of-concept implementations for all four. ArXiv · AI/CL/LG's note
score 4