LotusFA: A Federated Analytics System for Federated Learning of Watermarked Data

Federated Learning (FL) enables edge devices to collaboratively train global models while keeping data local, ensuring privacy preservation. However, the use of personalized watermarks for data ownership introduces the shortcut learning problem, which degrades model accuracy and training efficiency. We present LotusFA, a Federated Analytics (FA) system designed to mitigate these issues by i) locally estimating the watermark characteristics of each client's dataset; ii) collaboratively analyzing and adapting regularization parameters among clients to align with their watermark characteristics, thereby preventing over-reliance on watermark features, iii) integrating LotusFA as a library with the Flower framework to support edge devices. Deployed on 40 edge devices, LotusFA demonstrates real-world watermarked chest X-ray pneumonia analysis. LotusFA effectively balances model integrity and efficiency, offering a robust solution for watermark-aware federated learning.