Federated Benchmarking for Privacy-Preserving Urban Noise Monitoring

Abstract

This paper benchmarks a federated learning setup for acoustic event classification where raw recordings remain local to collection nodes. The authors evaluate aggregation rounds, client-dropout tolerance, and class-balance strategies using a shared protocol across five independently curated datasets. The best configuration improved macro-F1 by 9.6 points over a non-adaptive baseline while preserving data locality. Results highlight practical controls for robust privacy-preserving environmental audio analytics.