Source-linked AI summary
FedHome: Cloud-Edge based Personalized Federated Learning for In-Home Health Monitoring
Qiong Wu, Xu Chen, Zhi Zhou, Junshan Zhang
TL;DR
In-home health monitoring needs machine learning over abundant but privacy-sensitive data distributed across homes, while user data is imbalanced and non-IID. FedHome combines cloud-edge personalized federated learning with GCAE-based class balancing and reports stronger performance than established centralized and federated methods. The evaluation uses realistic human activity data traces.
Problem
Privacy-sensitive health data is distributed across homes, and existing federated approaches lack personalization while facing imbalanced and non-IID user data.
Method
FedHome trains a shared cloud model from local edge updates, keeps user data locally, and uses GCAE-generated balanced data to personalize local models.
Results
FedHome outperforms traditional centralized methods in balanced and imbalanced cases and reports 95.41% accuracy, over 7.49% above CNN and over 10% above other federated approaches.
Takeaways & Limitations
Cloud-edge federated learning with GCAE supports privacy-preserving, personalized in-home health monitoring while reducing model-transfer cost.
Takeaways & Limitations
The evaluation uses a specific realistic human activity dataset, and user data remains non-IID because users differ in characteristics, lifestyles, and medical histories.
Abstract
from arXiv · showhide
In-home health monitoring has attracted great attention for the ageing population worldwide. With the abundant user health data accessed by Internet of Things (IoT) devices and recent development in machine learning, smart healthcare has seen many successful stories. However, existing approaches for in-home health monitoring do not pay sufficient attention to user data privacy and thus are far from being ready for large-scale practical deployment. In this paper, we propose FedHome, a novel cloud-edge based federated learning framework for in-home health monitoring, which learns a shared global model in the cloud from multiple homes at the network edges and achieves data privacy protection by keeping user data locally. To cope with the imbalanced and non-IID distribution inherent in user's monitoring data, we design a generative convolutional autoencoder (GCAE), which aims to achieve accurate and personalized health monitoring by refining the model with a generated class-balanced dataset from user's personal data. Besides, GCAE is lightweight to transfer between the cloud and edges, which is useful to reduce the communication cost of federated learning in FedHome. Extensive experiments based on realistic human activity recognition data traces corroborate that FedHome significantly outperforms existing widely-adopted methods.
1 INTRODUCTION
FedHome addresses privacy, personalization, statistical heterogeneity, and communication challenges in in-home health monitoring through cloud-edge federated learning. Its GCAE-based personalization and experiments on realistic activity data are reported to outperform established methods.
- Motivation: Privacy-sensitive health data makes centralized collection risky and difficult to adopt for in-home monitoring.The paper also notes that health data scattered across families is difficult to integrate for model training.
- Challenges: Existing federated learning commonly trains one global model, which lacks personalization and struggles with users’ imbalanced and non-IID data.Communication overload within each federated round is also identified as an important efficiency challenge.
- Approach: FedHome combines cloud-edge federated learning with local data retention to train a shared model while supporting personalized learning.The framework coordinates dispersed homes through a central cloud server without uploading user data.
- Approach: GCAE synthesizes minority-class samples into a balanced dataset for local retraining, while parameter sharing and dimension reduction reduce transfer cost.The generated data supports personalization against imbalanced and non-IID monitoring distributions.
- Results: 95.41% accuracy is reported for FedHome, with more than 7.49% improvement over CNN and more than 10% gain over other federated approaches.Experiments use a realistic human activity dataset and compare balanced and imbalanced data cases.
2 OVERVIEW OF FEDHOME
FedHome uses a cloud-edge architecture and federated learning to process health data across homes while preserving local data custody. The overview emphasizes personalization, family or home-edge computation, and coordinated model updates.
- System setting: The system comprises a cloud server and dispersed edge nodes, such as smart-home gateways, for in-home monitoring tasks.The stated example is elderly fall-down detection via human activity recognition.
- Architecture: FedHome avoids centralized health-data uploading by using federated learning for collaborative training across homes.The architecture is motivated by privacy concerns associated with conventional centralized training.
- Architecture: Cloud-edge processing also addresses network latency and possible Internet interruption in emergency monitoring situations.The paper specifically mentions fall detection for elderly people as a real-time-sensitive case.
- Federated learning: FedHome coordinates edge computation and cloud aggregation by initializing client models from a common cloud-model state and exchanging updated parameters.The overview identifies federated learning as a multi-stage procedure involving local computation and cloud aggregation.
- Personalization: Personalization is needed because generic cloud models may not capture individual characteristics, while local personal data is often insufficient and highly skewed.The framework can also support family-shared models among mutually trusting household members.
3 LEARNING ALGORITHM FOR FEDHOME
FedHome combines cloud-edge federated learning with personalized GCAE training. The framework keeps health data at homes, addresses non-IID and imbalanced data through low-dimensional synthesis, and supports efficient model transfer.
- Personalized Federated Learning: Federated learning trains a shared global model from local client updates without sharing private user data.FedHome adopts FedAvg for cloud-edge collaborative training and adds homomorphic encryption to parameter exchange and gradient aggregation.
- Personalized Federated Learning: FedHome personalizes the cloud model by combining it with each home’s local health-monitoring data.The cloud model learns from dispersed homes and may not capture an individual user’s characteristics, motivating edge-side personalization.
- Generative Convolutional Autoencoder: GCAE jointly learns low-dimensional representations, reconstructs inputs, and predicts class labels using an encoder, decoder, and MLP.Its end-to-end objective combines prediction loss with decoder reconstruction loss, weighted by λ.
- Generative Convolutional Autoencoder: GCAE generates minority-class samples with SMOTE in low-dimensional space, producing a class-balanced dataset for refining personalized models.The encoder reduces 1200-dimensional inputs to 200 dimensions in the reported experimental setting, a reduction ratio of 6.
- Personalized Federated Learning: FedHome supports incremental learning by allowing cloud and edge models to continue adapting as new user data emerges.Each home can synthesize class-balanced data from personal observations and refine its model parameters.
4 EXPERIMENTS
The experiments evaluate FedHome on realistic human activity recognition data under balanced, imbalanced, and home-based partitions, comparing it with centralized and federated baselines. FedHome achieves strong accuracy while addressing non-IID data and communication efficiency through personalization and a compact GCAE model.
- Dataset and setup: The evaluation uses the MobiAct human activity recognition dataset, with 30 randomly selected volunteers treated as users in dispersed homes.Each user can train locally or offload computation to a trusted edge node.
- Dataset and setup: The study evaluates balanced, imbalanced, and home data partitions to represent different degrees of user and household data heterogeneity.Home partition groups 30 users into 10 homes with 1 to 5 family members who may share data at the edge.
- Performance evaluation: FedHome reaches 95.87% accuracy on balanced data and 95.41% on imbalanced data, while removing personalization lowers accuracy by 6.74% and 11.19%, respectively.These results are reported relative to the personalized FedHome framework.
- Communication efficiency: FedHome achieves 52,149 model parameters, whereas FL-MLP has nearly 30 times more; FL-CNN has fewer parameters but suffers 10.8–12.5% accuracy degradation.FedHome also significantly outperforms the larger FL-CNN-Large model in balanced and imbalanced partitions.
- Per-activity evaluation: FedHome achieves precision scores above 99% for JUM and STD activities, while precision and recall exceed 90% for fall and fall-like activities.The evaluation reports average precision, recall, and F1-score across all 30 users for each activity.
- Home data partition: For home data partition, FL-CNN produces user accuracies ranging from below 70% to above 95%, while FedHome-p improves prediction by combining autoencoder representations with parameter sharing.FL-MLP is more competitive but has large training and communication costs.
5 RELATED WORK
Related work covers sensor-based in-home health monitoring, cloud-edge healthcare systems, and federated learning. The paper positions FedHome as addressing privacy and federated-learning challenges in personalized monitoring.
- In-home health monitoring: In-home health monitoring uses mobile and ambient sensors, including smartphones and smartwatches, to improve healthcare effectiveness and reduce costs.Prior work includes continuous monitoring and automatic classification of Parkinson’s disease tremor severity with wearable accelerometer and gyroscope sensors.
- In-home health monitoring: UbeHealth proposes a personalized cloud-edge healthcare network, but its privacy treatment emphasizes organizational security policies rather than algorithm-level user privacy protection.The comparison distinguishes system-level security guidance from direct protection of user data during learning.
- Federated learning: Federated learning trains models over distributed devices without sharing raw datasets, while FedAvg provides a foundational aggregation framework for such settings.The related work also notes federated learning challenges including imbalanced and highly skewed data distributions.
6 CONCLUSION
FedHome combines cloud-edge federated learning with personalized health monitoring while keeping user data local. Its GCAE addresses imbalanced, non-IID data and reduces communication overhead, with experiments demonstrating effectiveness in performance and communication efficiency.
- FedHome trains a global model across multiple homes while protecting privacy by keeping user data locally.
- GCAE synthesizes minority-class samples to create a class-balanced dataset for personalized model learning.This targets the imbalanced and non-IID data dilemma in federated learning.
- GCAE contains few model parameters, significantly reducing communication overhead during cloud-edge model transfer.
- Extensive human activity recognition experiments demonstrate FedHome’s effectiveness in evaluation performance and communication efficiency.