Distributed learning comes with many advantages pertaining to network efficiency and data privacy. However, it also suffers from a variety of well-known problems, including non-iid distribution of data or labels, varying network latency and reliability, unequal participation, and agent failures. In this paper, we focus on the extreme case of unequal participant, which is known as persistent failure. We present two baseline heuristics that do not try to adapt to the agent failure, and which show that taking no action leads to significant loss in performance. We also introduce a proof-of-concept adaptive patching strategy, which, via model-inversion attack, extracts data points resembling the original training data used by the lost agent. After the attack, it restores the lost agent and supplies them with the extracted data points for further optimization. Using this approach, we show that it is possible to restore a sizeable amount of performance after an agent is lost, opening a new avenue for further research in patching strategies for persistent agent failures.
In decentralized learning, agents use their local data siloes to train their models, while regularizing to the neighbors they have access to. They all collaborate to reach a common goal of attaining the highest performance possible, while converging to the same model.
Imagine that one of the agents participating in the distributed optimization has been permanently dropped from the system. In scenarios like this, whenever that agent had some unique knowledge about the data distribution, the performance of the system can drop significantly.
So what can we do about it?
We propose a method that can restore the lost agent by attacking the latest model that we received from them. We want to perform a model-inversion attack to extract data points, which similar to those used by the lost agent during training. With these reconstructed data points, we can create a new agent and redeploy it to continue participating in the optimization process.
Wine
Iris
Digits
An analysis of the convergence of different decentralized learning algorithms and patching methods shows that the proposed model-inversion agent reconstruction method can restore the system's performance after a persistent agent failure and still lead to faster convergence than the baselines.
Wine
Iris
Digits
@article{Stepka2024strategies,
author = {Ignacy Stępka and Nicholas Gisolfi and Artur Dubrawski},
title = {Adaptive Strategies for Overcoming Persistent Agent Failures in Distributed Learning},
journal = {Robotics Institute Summer Scholars 2024},
year = {2024},
note = {Carnegie Mellon University},
}