# FLeet: Online Federated Learning via Staleness Awareness and Performance Prediction

**Authors:** Georgios Damaskinos, Rachid Guerraoui, Anne-Marie Kermarrec, Vlad Nitu, Rhicheek Patra, Francois Taiani  
**arXiv:** [2006.07273](https://arxiv.org/abs/2006.07273)  
**Date:** 2020-06-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2006.07273v1/

## Key points

- Federated Learning (FL) is a machine learning approach that preserves user data privacy while training a global model using updates computed on mobile devices.
- Standard FL infrastructures are not suitable for applications requiring frequent online model updates due to their lack of energy and performance efficiency.
- FLeet is the first Online FL system that acts as middleware between Android OS and the machine learning application, combining privacy benefits with precision of online learning through two core components: I-Prof and AdaSGD.
- I-Prof is a lightweight profiler that predicts and controls the impact of learning tasks on mobile devices, improving prediction accuracy up to 3.6x in computation time and up to 19x in energy consumption.
- AdaSGD is an adaptive learning algorithm resilient to delayed updates which outperforms alternative FL approaches by 18.4% in terms of convergence speed on heterogeneous data.
- Extensive evaluations showed that FLeet can deliver a 2.3x quality boost compared to Standard FL while consuming only 0.036% of battery per day.
- FLeet provides an efficient solution for online model updates while preserving user privacy and device performance, representing an important step forward in addressing the limitations of standard FL infrastructures.

## Summary

Federated Learning (FL) is a machine learning approach that allows for the training of a global model using updates computed on mobile devices while preserving user data privacy. However, standard FL infrastructures are not suitable for applications that require frequent online model updates due to their lack of energy and performance efficiency. To address this limitation, researchers have developed FLeet - the first Online FL system that acts as a middleware between the Android OS and the machine learning application. FLeet combines the privacy benefits of Standard FL with the precision of online learning through two core components: I-Prof and AdaSGD. I-Prof is a lightweight profiler that predicts and controls the impact of learning tasks on mobile devices, improving prediction accuracy up to 3.6x in computation time and up to 19x in energy consumption. AdaSGD is an adaptive learning algorithm resilient to delayed updates which outperforms alternative FL approaches by 18.4% in terms of convergence speed on heterogeneous data. Extensive evaluations showed that FLeet can deliver a 2.3x quality boost compared to Standard FL while consuming only 0.036% of battery per day. The results demonstrate how Online FL implemented by FLeet can provide an effective solution for applications requiring frequent online model updates without compromising privacy or device performance. Overall, FLeet represents an important step forward in addressing the limitations of standard FL infrastructures by providing an efficient solution for online model updates while preserving user privacy and device performance.