[论文解读] Semi-supervised Federated Learning for Activity Recognition

Yuchen Zhao, Hanyang Liu+3 more|arXiv (Cornell University)|Nov 2, 2020

Privacy-Preserving Technologies in Data参考文献 47被引用 35

一句话总结

本论文提出一种半监督联邦学习框架，用于在物联网边缘设备上的人体活动识别，使用本地自编码器从未标记数据学习表示，并在云端使用带标记表示的监督分类器。

ABSTRACT

Training deep learning models on in-home IoT sensory data is commonly used to recognise human activities. Recently, federated learning systems that use edge devices as clients to support local human activity recognition have emerged as a new paradigm to combine local (individual-level) and global (group-level) models. This approach provides better scalability and generalisability and also offers better privacy compared with the traditional centralised analysis and learning models. The assumption behind federated learning, however, relies on supervised learning on clients. This requires a large volume of labelled data, which is difficult to collect in uncontrolled IoT environments such as remote in-home monitoring. In this paper, we propose an activity recognition system that uses semi-supervised federated learning, wherein clients conduct unsupervised learning on autoencoders with unlabelled local data to learn general representations, and a cloud server conducts supervised learning on an activity classifier with labelled data. Our experimental results show that using a long short-term memory autoencoder and a Softmax classifier, the accuracy of our proposed system is higher than that of both centralised systems and semi-supervised federated learning using data augmentation. The accuracy is also comparable to that of supervised federated learning systems. Meanwhile, we demonstrate that our system can reduce the number of needed labels and the size of local models, and has faster local activity recognition speed than supervised federated learning does.

研究动机与目标

• Motivate HAR on edge IoT with privacy-preserving learning.
• Address lack of labeled data on edge devices in federated settings.
• Develop a semi-supervised FL pipeline using autoencoders to learn representations from unlabeled data.
• Evaluate how representation size and labeled data on the server affect performance.
• Demonstrate feasibility and efficiency on low-cost edge hardware like Raspberry Pi.

提出的方法

• Clients train autoencoders locally on unlabeled time-series data to learn representations.
• Server aggregates local autoencoders with FedAvg to form a global autoencoder.
• Server uses the encoder from the global autoencoder to transform labeled data into representations for supervised training of a classifier.
• Evaluate three autoencoder schemes: simple autoencoder (FC), convolutional autoencoder (CNN), and LSTM-autoencoder (LSTM-FC).
• Compare against baselines: centralised supervised learning (CS) and data augmentation-based semi-supervised FL (DA).
• Assess impact of label ratio and representation compression ratio on performance.
• Deploy and test the system’s local inference on a Raspberry Pi 4 Model B.

实验结果

研究问题

• RQ1 Q1: How does semi-supervised FL with autoencoders compare to centralized supervised learning?
• RQ2 Q2: How does semi-supervised FL with autoencoders compare to semi-supervised FL using data augmentation?
• RQ3 Q3: How does semi-supervised FL with autoencoders compare to supervised FL?
• RQ4 Q4: How do server label size and representation size affect HAR performance?
• RQ5 Q5: How efficient is semi-supervised FL on low-cost edge devices?

主要发现

• Semi-supervised FL with LSTM autoencoders and a Softmax classifier achieves higher accuracy than a centralized supervised system.
• The proposed system outperforms data augmentation–based semi-supervised FL (DA).
• The method attains comparable accuracy to supervised FL while using fewer labels and smaller local models.
• Using autoencoders reduces local model size and can lower upload traffic to the server.
• Inference on a Raspberry Pi 4 Model B is feasible with acceptable real-time processing times, compared to supervised FL.