Title: Federated Learning with Differential Privacy based Data Protection Model for Enhancing Utility and Convergence
Presenter: Dr. GUPTA Rishabh
Abstract: Federated learning (FL) is an emerging approach for training machine learning models on a central server in the cloud environment.
Unlike existing works that gather and train models on the central server in the cloud, FL enables various users to collaboratively train a global model on the central server
without disclosing their local data to the cloud. However, FL does not operate more effectively when each user only has a limited or subset of data. Thus, it suffers from
the issues of convergence and accuracy degradation problems. This paper proposes a novel Federated Learning with Differential Privacy based Data Protection Model (FD-DPM) to
enhance utility and convergence in the cloud environment. In the proposed model, the Laplace mechanism is considered to preserve the privacy of users’ data before transferring
it to the cloud. The global model is trained using user-provided data, the local model is trained using the local dataset, and computed values are transferred rather than the raw
data to the central server. Users acquire an updated average parameter based on averaging various local and global versions. This updated parameter is repeatedly shared with the
user at regular intervals so they can be retrained to develop a more robust, effective model that can accurately anticipate outcomes.
Download
|