A Multi-Trapdoor Editable Blockchain Scheme Based on Shamir Secret Sharing for Federated Learning

Abstract

Federated learning is a popular learning mechanism. A server coordinates many clients to complete the training process of the model. However, there is a problem of poisoning attacks in federated learning. For example, the client may send malicious parameters to damage the global model's performance. To reduce the clients' dishonest behavior during training, some schemes are committed to scoring the clients' parameters to get the reputation value and storing the reputation value on the blockchain to achieve the clients' incentive and further reduce the poisoning attack during the federated learning process.

However, due to the non-editable nature of the blockchain, when the reputation value stored on the blockchain is incorrect, it cannot be modified. This problem leads to the unfairness of the scheme that relies on traditional blockchain to store reputation value to motivate clients. We improved the original chameleon hash function to solve this problem and proposed a new multi-trap editable blockchain scheme.

Besides, we combined the dynamic Shamir secret sharing technology to ensure the trap door's security and avoid trap door centralization. We use our editable blockchain to store the reputation value generated in federated learning. It not only realizes the client's incentive but also effectively reduces dishonest behavior during training and supports the modification of the reputation value on the blockchain.

Our scheme ensures the client can participate fairly in the federated learning training task. Experiments show that our editable blockchain scheme supports the safe modification of reputation value and has significant function advantages compared with existing schemes.