FHDnn: Communication Efficient and Robust Federated Learning for AIoT Networks
TimeTuesday, July 12th11:15am - 11:37am PDT
Location3000, Level 3
Event Type
Research Manuscript
AI/ML Security/Privacy
ML Algorithms and Applications
DescriptionFederated learning is a popular strategy for learning models in a distributed fashion for IoT applications. However, communication resources, unreliable network and compute resources are major bottlenecks for federated learning.
We propose FHDnn, a novel federated learning framework that is robust to network errors. FHDnn combines CNNs with HD computing and accelerates learning by training only the HD model. FHDnn leverages online learning capabilities of HD computing to improve accuracy. We show that FHDnn reduces communication costs and local computations. FHDnn is also highly robust to poor network conditions such as packet loss and noise with minimal loss in accuracy.