Federated Learning with Noisy User Feedback

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2726 - 2739 July 10-15, 2022 ©2022 Association for Computational Linguistics Federated Learning with Noisy User Feedback Rahul Sharma∗ rahul.sharma@usc.edu Anil Ramakrishna∗ aniramak@amazon.com Ansel MacLaughlin ammaclau@amazon.com Anna Rumshisky arum@cs.uml.edu Jimit Majmudar mjimit@amazon.com Clement Chung chungcle@amazon.com Salman Avestimehr avestime@usc.edu Rahul Gupta gupra@amazon.com Abstract Machine Learning (ML) systems are getting increasingly popular, and drive more and more applications and services in our daily life. This has led to growing concerns over user privacy, since human interaction data typically needs to be transmitted to the cloud in order to train and improve such systems. Federated learn- ing (FL) has recently emerged as a method for training ML models on edge devices using sen- sitive user data and is seen as a way to mitigate concerns over data privacy. However, since ML models are most commonly trained with label supervision, we need a way to extract labels on edge to make FL viable. In this work, we propose a strategy for training FL models us- ing positive and negative user feedback. We also design a novel framework to study differ- ent noise patterns in user feedback, and explore how well standard noise-robust objectives can help mitigate this noise when training models in a federated setting. We evaluate our proposed training setup through detailed experiments on two text classification datasets and analyze the effects of varying levels of user reliability and feedback noise on model performance. We show that our method improves substantially over a self-training baseline, achieving perfor- mance closer to models trained with full super- vision. 1 Introduction Artificial Intelligence (AI) and Machine Learning (ML) have become increasingly common in mod- ern society with applications ranging from simple standalone products to complex modules Kaplan and Haenlein (2019). However, this rise has also created growing privacy concerns Papernot et al. (2016); Yeom et al. (2018). Such concerns may affect user willingness the adapt new technologies Guhr et al. (2020). In response, many government agencies across the world have introduced regu- lations to protect the data-handling rights of their ∗equal contributions citizens, such as the European Union’s GDPR Gen- eral Data Protection Regulation and California’s CCPA California Consumer Privacy Act. Federated Learning (FL) is a step in this direc- tion to improve consumer trust, where models are trained without moving data out of client devices. The typical FL approach is to iteratively train local models on edge devices and then propagate them back to a central node in order to update the global model. Most commonly, this is done using Feder- ated Averaging (FedAvg) McMahan et al. (2017), where we take a simple average over the client model parameters. However, in order to update local models on the edge, this setup assumes the presence of labeled user data on each device, which is often not possible. Most prior works do not ad- dress this problem, but simulate fully-supervised federated learning by distributing existing labeled datasets …

Content truncated to 500 words