ImprovDML: Improved Trade-off in Private Byzantine-Resilient Distributed Machine Learning

Bing Liu, Chengcheng Zhao, Li Chai, Peng Cheng, Yaonan WangarXiv 2025 · DOI · arXiv

Introduction

In the evolving landscape of Distributed Machine Learning (DML), collaborative training across decentralized devices—such as smartphones, autonomous vehicles, or industrial robots—has become a cornerstone of modern AI. However, this decentralization introduces two critical vulnerabilities: Byzantine attacks, where malicious participants sabotage the model with faulty data, and privacy attacks, where sensitive local data is reconstructed from shared updates.

The Challenge of Decentralized Learning

Traditional DML often relies on a central server to aggregate updates, known as Server-worker DML. While easier to manage, this creates a single point of failure and a potential privacy bottleneck. Decentralized DML removes the central server, allowing nodes to communicate directly in a peer-to-peer fashion. However, this shift makes the system more susceptible to adversarial influence because there is no trusted central authority to filter updates.

Comparison of Server-worker and Decentralized DML architectures.

Existing defenses typically combine Byzantine-resilient aggregation with Differential Privacy (DP). Unfortunately, this "double defense" often leads to a significant drop in model accuracy. Standard aggregation rules frequently produce estimates that fall outside the "safe zone" of honest data, and the heavy noise required by traditional DP can drown out the learning signal entirely, rendering the model useless.

The ImprovDML Framework

To address these trade-offs, the authors propose ImprovDML, a framework that integrates two primary innovations: Resilient Vector Consensus (RVC) and Concentrated Geo-Privacy (CGP).

Resilient Vector Consensus (RVC)

Most Byzantine defenses attempt to estimate the mean of honest updates, which often results in an aggregation point outside the optimal range. ImprovDML utilizes RVC algorithms that are mathematically guaranteed to find an aggregation point strictly within the convex hull of honest agents.

Geometric representation of the convex hull advantage. The green star represents the RVC aggregation point within the safe zone (grey), avoiding the error (yellow star) typical of standard methods.

By staying within this hull, ImprovDML eliminates the "non-consensus" error that typically plagues decentralized learning, ensuring that the global model remains representative of legitimate data even in the presence of malicious actors.

Concentrated Geo-Privacy (CGP)

Rather than applying traditional $(\epsilon, \delta)$-Differential Privacy, which treats all input changes as equally sensitive, ImprovDML adopts Concentrated Geo-Privacy (CGP). CGP accounts for the Euclidean distance between inputs, allowing for a more "fine-grained" privacy guarantee. This approach requires significantly less noise to achieve the same level of security, preserving the utility of the training updates and leading to higher model accuracy.

Experimental Evaluation

The researchers tested ImprovDML in a multi-robot learning scenario using a binary classification task on synthetic XOR data. The setup involved 14 robots (13 honest, 1 Byzantine) with training sets ranging from 1,100 to 1,600 samples per node. The resulting algorithm, PP-RSGD (Private Byzantine-Resilient Stochastic Gradient Descent), follows a cycle of local training with noise injection, peer-to-peer transmission, and RVC-based aggregation.

Key Results and Insights

  • Superior Accuracy: ImprovDML consistently outperformed state-of-the-art frameworks like PP-IOS in test accuracy across simulations.
  • Faster Convergence: The "consensus error"—the variance between individual robot models—decreased more rapidly and remained lower than in previous methods, indicating a more stable learning process.
  • Optimized Privacy Trade-off: By using CGP, the authors demonstrated that privacy parameters could be orders of magnitude tighter than standard DP for the same noise level ($\sigma = 2.0$).
  • Resilience to Inversion: The framework successfully defeated Gradient Inversion Attacks, proving that the added noise effectively masked private training data without compromising the learning process.

Conclusion and Future Directions

ImprovDML represents a significant step forward in balancing security and utility in decentralized machine learning. While the computational complexity of RVC algorithms remains a challenge for high-dimensional deep learning models, the use of linear programming solvers offers a path forward. Future work may explore temporally varying noise levels and more complex network topologies to further refine the efficiency of private, resilient distributed training.

Summary updated

BibTeX
@misc{liu2025improvdml,
  author = {Bing Liu and Chengcheng Zhao and Li Chai and Peng Cheng and Yaonan Wang},
  title = {{ImprovDML: Improved Trade-off in Private Byzantine-Resilient Distributed Machine Learning}},
  year = {2025},
  eprint = {2506.15181},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.15181},
}
Figure