[Paper Review] Prophet: Proactive Candidate-Selection for Federated Learning by Predicting the Qualities of Training and Reporting Phases

Huawei Huang, Kangying Lin+3 more|arXiv (Cornell University)|Feb 3, 2020

Privacy-Preserving Technologies in Data16 references4 citations

TL;DR

This paper proposes Prophet, a proactive candidate-selection framework for Federated Learning (FL) that uses LSTM to predict device quality during training and reporting phases, and employs Deep Reinforcement Learning (DRL) to proactively select reliable participants. Evaluated on real-world traces, Prophet significantly reduces FL latency and improves success rates compared to reactive selection methods.

ABSTRACT

Although the challenge of the device connection is much relieved in 5G networks, the training latency is still an obstacle preventing Federated Learning (FL) from being largely adopted. One of the most fundamental problems that lead to large latency is the bad candidate-selection for FL. In the dynamic environment, the mobile devices selected by the existing reactive candidate-selection algorithms very possibly fail to complete the training and reporting phases of FL, because the FL parameter server only knows the currently-observed resources of all candidates. To this end, we study the proactive candidate-selection for FL in this paper. We first let each candidate device predict the qualities of both its training and reporting phases locally using LSTM. Then, the proposed candidateselection algorithm is implemented by the Deep Reinforcement Learning (DRL) framework. Finally, the real-world trace-driven experiments prove that the proposed approach outperforms the existing reactive algorithms

Motivation & Objective

• To address the high training latency in Federated Learning caused by unreliable device participation.
• To overcome the limitations of reactive candidate-selection methods that rely only on current device resource observations.
• To design a proactive system that predicts future device performance during both training and reporting phases.
• To improve FL efficiency and convergence by selecting only high-quality, reliable devices in advance.

Proposed method

• Each client device locally predicts the quality of its training and reporting phases using Long Short-Term Memory (LSTM) networks.
• The system employs a Deep Reinforcement Learning (DRL) framework to make proactive candidate-selection decisions based on predicted device qualities.
• The DRL agent learns to select devices that are likely to successfully complete both training and reporting phases.
• The framework uses real-world device trace data to train and validate the prediction and selection models.
• The prediction and selection pipeline is designed to minimize FL latency by avoiding devices with high failure risk.
• The approach shifts from reactive to proactive selection by anticipating device reliability before the FL round begins.

Experimental results

Research questions

• RQ1 Can LSTM-based local prediction of training and reporting phase quality improve FL candidate selection?
• RQ2 How does proactive selection using predicted device quality compare to reactive selection based on current observations?
• RQ3 To what extent can Deep Reinforcement Learning optimize candidate selection under dynamic device availability?
• RQ4 What is the impact of proactive selection on FL convergence latency and success rate in real-world scenarios?

Key findings

• Prophet significantly reduces training latency in Federated Learning by proactively avoiding devices likely to fail during training or reporting.
• The use of LSTM for local quality prediction enables more accurate forecasting of device reliability than reactive methods.
• The DRL-based selection framework outperforms existing reactive algorithms in terms of success rate and convergence stability.
• Real-world trace-driven experiments confirm that Prophet improves overall FL efficiency and reliability.
• The proactive approach leads to higher completion rates of FL rounds compared to reactive selection strategies.
• The system demonstrates robust performance under dynamic network and device conditions, validating its practical applicability.