Trustworthy Federated Learning and Model Governance

Abstract: Federated learning is an essential intersection of artificial intelligence and private computing. How to make federated learning more secure, reliable, and efficient is the future focus of industry and academia. This report will systematically review the progress and challenges of federated learning and look ahead to several significant developments.
 

Copyright Validation for Federated Deep Learning

Abstract: During the development of federated models, there are risks of illegal copying, redistribution, and abuse. To address these risks, we propose a novel Federated Deep Neural Network (FedDNN) ownership verification scheme that allows private watermarks to be embedded and verified to claim the legitimate IPR of the FedDNN model.
 

AI for Ciphertext Semantic Retrieval and Crypto for AI Copyright Protection

Abstract: This talk illustrates the use of AI technology to realize the ciphertext semantic retrieval, and the cryptography methods to better implement AI copyright supervision. Searchable encryption schemes have been studied for many years, but most of them can only achieve exact matching between keywords and query terms at present. Some works have achieved ciphertext semantic retrieval, however, these works can only perform semantic extension the query in the plaintext domain and cannot achieve semantic matching in the ciphertext domain. This talk discusses how to solve this problem using word vectors generated by AI and provides an outlook for future research on ciphertext semantic retrieval. In addition, machine learning models are regarded as important digital products and assets, but the AI model copyright supervision is mainly oriented to traditional DNN models. However, there is still seldom works on copyright protection in the context of federated learning. Therefore, this talk summarizes the challenges encountered in model copyright protection in federated learning, and through the perspective of cryptography, making use of related theories to further realize Copyright protection under Federated Learning. This talk also provides a novel watermarking scheme to copyright-protect federated learning models, and a novel federated learning model traceability framework named FedTracker.