Design Considerations and Guidelines
Although research into model governance for FedL is still relatively immature , opportunities exist with Blockchain to ensure trust in the models being deployed for AI based systems , especially those shared , licensed , or purchased from third parties . In the context of FedL , if a client is compromised to enter incorrect model update ( model poisoning attack ( Chen , 2021 )), Blockchain provides a consensus mechanism about the quality of the model update received from that client i . e ., whether to accept the model update or reject .
One such example is proposed by ( Zhang , 2021 ) which utilizes blockchain to build a data-sharing sharing platform for manufacturing organizations . The organizations are categorized into : ( 1 ) client organizations and ( 2 ) server organizations . Clients train the federated model while the servers owned the data-sharing platform and orchestrated the model aggregation process .
Another example of utilizing blockchain for decentralized model aggregation is proposed in ( Zhao , 2021 ) where models related to home appliances were used to fine tune their performance i . e ., energy consumption . Blockchain is again used as a model sharing platform where consumers can upload their specific local models to be aggregated by dedicated miners on the blockchain .
Given the advantages , Blockchain also faces some challenges which need to be addressed before utilizing it for FedL . In such de-centralized setup , clients must share their model updates with each other to reach to the consensus about the optimality of the global model . This may expose an organization ’ s local model parameters to another competing organization .
Thus , another layer of security at client level is needed on top of DLT to protect the sovereignty of the model updates . Solutions such as differential privacy ( DP ) ( Choudhary , 2019 ) and homomorphic encryption ( Shreshtha , 2019 ) can prevent the re-construction of raw data from client model update . However , both are time consuming , and DP suffers in terms of model accuracy . They can be utilized along with Blockchain for cross-silo FedL whereas , they are not suitable for cross-device FedL where real-time data analytics is of paramount importance and thus a trusted centralized and / or hybrid orchestration should be an appropriate choice for such scenarios .
Once the appropriate client , mode of operation , security provisions and orchestration type are selected for the given smart manufacturing use case requirements , suitable technological drivers can be derived according to the possible FedL solution configuration ( Table 4-1 ).
While the FedL implementation guidelines proposed in section 5 can be leveraged to support all use case examples as set out in section 2 , this section presents the use of the decision model shown in Figure 4-1 in the context of product optimization to identify the implementation guidelines .
30 March 2022