Unlinkable Collaborative Learning Transactions: Privacy-Awareness in Decentralized Approaches

Abstract

Smart contracts (SCs) and collaborative learning (CL) are disclosed publicly, in which most transactions and activities that occur by the parties can be bared in real-time. Both are strengthened in a decentralized manner. CL allows numerous clients to collectively build deep learning models privately by aggregating the gradient values from clients' devices, yet it lacks the incentive mechanism for the contributing clients. On the other hand, the merits of SCs can be a plausible solution as an incentive mechanism in the CL system because self-executing contracts with immutable data records are resistant to failure. The clients can claim the rewards by stating their contribution arbitrarily in the SCs and tendering a proof transaction function. Nevertheless, directly adopting SCs in the CL system could breach clients' privacy because the transactions are exposed openly. The observer can infer the properties of the clients' resources. Therefore, we designed schemes that can overcome observers' ability to link clients' information with their associated devices during training. In essence, our schemes are unbiased. We also provide a secure incentive mechanism for the parties in the CL system by obscuring the information values. Finally, the numerical results indicate that the proposed schemes satisfy the design goals.