1: \begin{abstract}

2: 

3: 

4: Federated Learning (FL) is very appealing for its privacy benefits: 

5: essentially, a global model is trained with updates computed on mobile devices while keeping the data of users local. 

6: Standard FL infrastructures are however designed to have no energy or performance impact on mobile devices, and are therefore not suitable for applications that require frequent (\emph{online}) model updates, such as news recommenders.%

7: 

8: This paper presents \system, the first \emph{Online FL} system, acting as a middleware between the Android OS and the machine learning application.

9: \system combines the privacy of Standard FL with the precision of online learning thanks to two core components: (i) \profiler, a new lightweight profiler that predicts and controls the impact of learning tasks on mobile devices, and

10: (ii) \algo, a new adaptive learning algorithm that is resilient to delayed updates. %

11: 

12: Our extensive evaluation shows that Online FL, as implemented by \system, can deliver a 2.3$\times$ quality boost compared to Standard FL, while only consuming 0.036\% of the battery per day.

13: \profiler can accurately control the impact of learning tasks by improving the prediction accuracy up to 3.6$\times$ (computation time) and up to 19$\times$ (energy). 

14: \algo outperforms alternative FL approaches by 18.4\% in terms of convergence speed on heterogeneous data.

15: 

16: 

17: \end{abstract}

18: