Robust and effective online Federated Learning – FLawed

Mobile internet use has already surpassed desktop use since several years ago, and the gap continues to widen every year. The large amounts of mobile generated data can be leveraged through different Machine Learning techniques for high quality and more personalized internet services. In this context, Federated Learning (FL) is very appealing for its privacy benefits: essentially, a global model is collaboratively trained right where the data is created, on mobile devices. Recent work have shown that the traditional FL architecture where the data is generally integrated in the model during the night is not suitable for applications that operate on very timely data (e.g. recommandation systems), leading to the onset of online FL. Despite its increasing popularity, online FL has several limitations which may slow down its widespread adoption. First, in online FL, the computation of model updates drain energy directly from the phone's battery. Therefore, it may be necessary to motivate users with incentives to contribute to the global objective. Second, mobile operating systems were designed for responsive UI applications so they lack the necessary abstractions for an optimal execution of FL tasks which can be classified as batch CPU-intensive tasks. Third, the online FL architecture (and FL in general) is sensitive to numerous robustness attacks because the users have not only access to the global model's parameters but can also influence their value through the model updates sent to the FL server. The objective of FLawed is to address the aforementioned issues and build a robust and effective online FL framework, viable in the context of real-life applications.