[Paper Review] Active Federated Learning

Jack Goetz, Kshitiz Malik+4 more|arXiv (Cornell University)|Sep 27, 2019

Privacy-Preserving Technologies in Data13 references83 citations

TL;DR

Active Federated Learning (AFL) selects client cohorts based on a value function computed from the current model and client data, reducing training iterations by 20-70% while preserving accuracy.

ABSTRACT

Federated Learning allows for population level models to be trained without centralizing client data by transmitting the global model to clients, calculating gradients locally, then averaging the gradients. Downloading models and uploading gradients uses the client's bandwidth, so minimizing these transmission costs is important. The data on each client is highly variable, so the benefit of training on different clients may differ dramatically. To exploit this we propose Active Federated Learning, where in each round clients are selected not uniformly at random, but with a probability conditioned on the current model and the data on the client to maximize efficiency. We propose a cheap, simple and intuitive sampling scheme which reduces the number of required training iterations by 20-70% while maintaining the same model accuracy, and which mimics well known resampling techniques under certain conditions.

Motivation & Objective

• Motivate reducing communication and computation costs in Federated Learning by smarter client selection.
• Introduce AFL as a framework that uses a value function evaluated on clients to guide sampling.
• Show that AFL can achieve the same accuracy with significantly fewer training rounds (20-70% fewer).
• Discuss practical considerations including privacy (differential privacy) and potential extensions.

Proposed method

• Define a value function V that evaluates client data on-device using the current model w^(t).
• Servers collect valuations {v_k} from selected clients and convert them into a sampling distribution for the next round.
• A natural choice is a loss-based valuation v_k = (1 / sqrt(n_k)) l(x_k, y_k; w^(t)).
• Convert valuations to probabilities via a softmax-like mechanism with tuning parameters (α1, α2, α3).
• Incorporate a differential privacy perspective and discuss privacy-preserving querying (e.g., Sparse Vector) to prevent leakage.
• Experimentally compare AFL to uniform random client selection on Reddit and Sticker Intent datasets, showing 20-70% fewer epochs for similar accuracy.

Experimental results

Research questions

• RQ1 Can active, model-informed client selection reduce the number of training rounds in Federated Learning without sacrificing accuracy?
• RQ2 How should valuations be computed on-device to best reflect a client's utility for model improvement?
• RQ3 What is the impact of differential privacy on valuation-based client selection and overall performance?
• RQ4 How does AFL compare to standard Random Sampling and server-side resampling in practical datasets?
• RQ5 What tuning parameters govern AFL’s sampling and how do they affect efficiency and accuracy?

Key findings

• AFL achieves similar model accuracy with 20-70% fewer training epochs compared to uniform client selection.
• Using loss-based valuations correlates with higher utility clients and accelerates convergence.
• AFL’s client-centric sampling significantly outperforms server-side data-point resampling in observed gains on Reddit data.
• AFL frameworks demonstrate practical feasibility with on-device valuation computations and privacy considerations.
• AFL’s gains depend on tuning parameters α1, α2, α3 and the valuation function used.