Workshop: FTXS: Workshop for Faults, Trustworthiness, and eXplainability for AI Systems at Scale
Authors: Ruiwen Shan and Jon C. Calhoun (Clemson University)
Abstract: Federated Learning(FL) trains models across clients without sharing local data, but its iterative training process requires frequent transmission of model updates and introduces high communication overhead. Error-bounded lossy compression can effectively reduce this overhead. However, existing studies mainly focus on data reduction and accuracy but have limited attention to communication reliability. Compressed bitstreams are sensitive to transmission errors, and a few corruptions can cause an entire model layer to fail reconstruction. We propose ARC-FL, a framework that combines hierarchical compression with unequal error protection. ARC-FL assigns model layers to three protection levels by parameter update magnitude and uses a layer-wise fallback for partial decoding failures. At a BER of 3E-3, equal error protection(EEP) with strong protection fails to maintain model convergence in several configurations. Meanwhile, ARC-FL preserves meaningful accuracy across all tested architectures and datasets and achieves up to 7.5x the accuracy of EEP in the most affected configuration.
Back to FTXS: Workshop for Faults, Trustworthiness, and eXplainability for AI Systems at Scale Archive Listing
Back to Full Workshop Archive Listing