Global Model Privacy Protection Mechanism in Federated Learning
Ajit Kumar, Bong Jun Choi · 2024
Data scarcity is a crucial concern in the traditional approach of training deep learning and one of the bottlenecks that limit its growth. Recently, Federated Learning (FL) has become a suitable approach for providing data privacy and is emerging as a solution for data scarcity. However, FL has opened up a new issue, i.e., model privacy and security. In vanilla FL, each participant receives the updated global model in every training round. Hence, if a model trainer wants to keep the updated global model private from participants, there is limited scope to protect the model access. There needs to be more literature on preserving the global model, and possible solutions like differential privacy, cryptography, or subnetworks are insufficient. In the proposed work, we have introduced the privacy issues in the global model and provided experimental results to demonstrate global model leaks, i.e., each participant has a model with equivalent accuracy to the global model in the subnetwork-based FL approach.