Time: 9.00 – 11.00 a.m., March 29th, Friday
Venue: Lecture Hall, Floor 9, College of Big Data, Mingxiang Campus
Speaker: Zhibo Wang
associate professor of Cyber Science and Engineering, Wuhan University
doctor degree of computer engineering at University of Tennessee in USA
bachelor degree of automation in Zhejiang University
Member of IEEE, ACM and CFF
Research Areas: Internet of Things, Mobile Awareness and Computing, Network Security and Privacy Protection
Over 70 papers relevant to network and security, including 16 CCF A-level papers, 4 ESI highly cited papers
Abstract: Horizontal Federated Learning is a distributed machine learning framework, which have received extensive attention and research in privacy security and machine learning these days. Compared with the traditional centralized learning framework, horizontal federated learning transfers the model training process to the user side, and only requires the user to submit the model parameter periodically to complete the model training, thus avoiding the malicious access and abuse of the user data by the server side. The report examines privacy issues in horizontal federal learning and proposes an attack method of user privacy data reconstruction based on malicious server. The user data distribution is simulated by building a multi-task Generative adversarial network model, and the user parameter update is used to calculate the data representation to reconstruct the specific user privacy data. Compared with the existing attack methods, which can only reconstruct the sample data representing a certain category, the attack method can realize the user-level data reconstruction. The effectiveness of the attack is verified by handwritten number classification and face recognition. The model parameter update contains too much privacy information, so the existing horizontal federal learning frameworks still have security risks.
From College of Big Data
Translated and Edited by Weiwei Wang
Go Back