CO-OP: Cooperative Machine Learning from Mobile Devices

Massive amounts of user behavior logs and sensor data are generated on mobile devices, which can help to improve the usability of social media apps and other intelligent apps. However, collecting such personal data may spark privacy and legal concerns. Recently, many efforts in both academia and industry have been devoted to developing distributed machine learning methods and architectures to scale up to a large amount of distributed data. However, most such methods focus on a server cluster or data residing on several datacenters. This work takes one more step toward the more ambitious objective of collectively training machine learning models based on data from mobile devices, yet without collecting such private data centrally. We propose CO-OP, an asynchronous protocol which leverages the computing power of each mobile client to train local models based on small amounts of newly generated training samples, and merges such local models into a global model on the server judiciously, balancing the model accuracy and communication overhead. We implemented a CO-OP Android app and tested it on 60 real clients distributed in different continents. Results suggest that CO-OP can achieve an accuracy of more than 80% on image classification using neural networks based on the MNIST datasets, even when the clients are intermittently available and training data are generated dynamically on the go. Additional simulation results also demonstrate the effectiveness of CO-OP in training Support Vector Machine (SVM) and Logistic Regression models.