Skip to content

Back to the catalog

DPF-SGD: Byzantine-Robust Differentially Private Federated SGD with Exact Optimality Guarantees
eprint.iacr.orgpaper

DPF-SGD: Byzantine-Robust Differentially Private Federated SGD with Exact Optimality Guarantees

DPF-SGD is a federated SGD algorithm that stays robust to Byzantine agents while satisfying ϵ-differential privacy. It combines private step sizes with a DP noise-decay schedule and adds a privacy-preserving comparative elimination (PPCE) filter that suppresses Byzantine updates during aggregation. The authors prove almost-sure convergence to the global optimum despite persistent noise and adversaries. On synthetic and MNIST data, final optimization error is one to two orders of magnitude lower than DP baselines.

TL;DR Byzantine-robust, ϵ-DP federated SGD with a PPCE filter and almost-sure convergence to the optimum, beating DP baselines by 10–100x on MNIST and synthetic data.
Open original ↗
Document

How to use it

There is no code, repository or artifact on the page. The paper is a preprint on the Cryptology ePrint Archive (Paper 2026/2467, category Applications) by Likang Lu and Jiaxin Wu. It was received and approved on 2026-10-14 and is released under CC BY, so the method and text can be reused with attribution. To reproduce the results you would have to implement DPF-SGD and the PPCE filter from the full PDF.

Limits

  • Only the abstract is available here. The page does not state any of the following:
  • the ε values tested
  • the fraction or behaviour of Byzantine agents assumed
  • the number of federated agents
  • the model trained on MNIST
  • which "DP baselines" were compared
  • the absolute error values behind the "one to two orders of magnitude" reduction

Without these, the headline result cannot be judged against your own privacy budget or threat model.

  • No conditions are given for the convergence guarantee. The abstract claims almost-sure convergence to the global optimum, which suggests restrictions such as convexity or bounds on how many agents are Byzantine. It names neither. Check these in the PDF before relying on the "exact optimality" claim for non-convex models.
  • Experiments are small-scale. They cover only synthetic data and MNIST. No results on larger datasets or on non-IID client splits are mentioned.
  • The PPCE filter's cost is not described. The abstract says the filter "preserves privacy guarantees", but it gives no computation or communication cost and no explanation of how a comparison-based filter avoids leaking information about honest updates.
  • The paper is not peer-reviewed. It is a same-day ePrint posting. Use "See all versions" on the page to check for revisions.

Technologies named here

  • MNIST
  • DPF-SGD
  • SGD

Techniques named here

  • DPF-SGD
  • Privacy-preserving comparative elimination filter

Related in research