Caring Without Sharing: A Federated Learning Crowdsensing Framework for Diversifying Representation of Cities
This work addresses privacy issues in data collection for AI models, offering a novel approach for domain-specific applications like assistive technology, though it appears incremental as it applies existing federated learning to a new context.
The paper tackles the privacy concerns in mobile crowdsensing by proposing a federated learning framework that enables collaborative model training without sharing personal data, demonstrated through a case study on diversifying vision algorithms for sidewalk obstacle representation to aid visually impaired navigation.
Mobile Crowdsensing has become main stream paradigm for researchers to collect behavioral data from citizens in large scales. This valuable data can be leveraged to create centralized repositories that can be used to train advanced Artificial Intelligent (AI) models for various services that benefit society in all aspects. Although decades of research has explored the viability of Mobile Crowdsensing in terms of incentives and many attempts have been made to reduce the participation barriers, the overshadowing privacy concerns regarding sharing personal data still remain. Recently a new pathway has emerged to enable to shift MCS paradigm towards a more privacy-preserving collaborative learning, namely Federated Learning. In this paper, we posit a first of its kind framework for this emerging paradigm. We demonstrate the functionalities of our framework through a case study of diversifying two vision algorithms through to learn the representation of ordinary sidewalk obstacles as part of enhancing visually impaired navigation.