Researchers from UII America have recently presented their latest work on privacy-preserving federated learning at the 2022 AAAI Conference on Artificial Intelligence (AAAI). AAAI is considered to be one of the top conferences in the field of artificial intelligence. In 2022, the conference has a record-low acceptance rate of 15%. The paper, entitled “Preserving Privacy in Federated Learning with Ensemble Cross-Domain Knowledge Distillation,” was among the 50 papers accepted by the “AI for Social Impact” technical track. In this work, the UII America research team focuses on generalizing the distillation-based federated learning framework to tasks such as tumor segmentation on MRI images and natural language processing in addition to natural image classification and segmentation. The proposed one-shot one-way federated learning framework (FedKD) distills knowledge with unlabeled, cross-domain non-proprietary data. Due to the proposed seminal quantized and noisy ensemble algorithm, which meaningfully minimizes privacy cost, FedKD is capable of preserving local data privacy without sacrificing accuracy at a substantially lower communication bandwidth. The research team demonstrates superior performance in terms of accuracy, bandwidth, and privacy guarantee compared to prior arts on various benchmark datasets.
Figure 1: Traditional federated learning methods transfer private parameters or gradients from local nodes to a server, risking privacy leakage. FedKD from UII America trains local models independently, and only transfers products of the unlabeled public data.
Follow UII on WeChat
© 2019 United Imaging Intelligence.
All Rights Reserved.