Power to the Clients: Federated Learning in a Dictatorship Setting

Mohammadsajad Alipour, Mohammad Mohammadi AmiriarXiv 2025 · DOI · arXiv

Introduction

Federated Learning (FL) is built on the promise of collaborative intelligence, where multiple clients contribute gradients to train a shared model without exposing raw data. While decentralized training offers privacy benefits, it also introduces significant security risks. Most existing research focuses on Byzantine attacks, which aim to degrade model performance, or Backdoor attacks, which insert hidden triggers.

This paper identifies a third, underexamined threat: Dictatorship. In this scenario, a malicious client (or group) is selfish rather than purely destructive. Their goal is to hijack the global model so it performs perfectly on their local data distribution while effectively "erasing" the contributions of all other participants. This turns a collaborative system into a private training session subsidized by the computation of others.

The Dictatorship Attack Mechanism

The core of the dictatorship attack relies on mathematical reconstruction to cancel out benign influence.

  1. Single Dictator Strategy: The dictator maintains a local "shadow model" representing the state of the model if only they were training. By comparing the server’s broadcasted global model with this shadow model, the dictator can isolate the aggregate influence of all other clients. They then submit a manipulated update designed to perfectly subtract that collective influence while injecting their own desired gradient.
  2. Collaborative Dictators: Malicious clients can coordinate via side channels to share gradients, acting as a unified "super-dictator." This allows them to outweigh benign participants even when individual update sizes are restricted.
  3. Learning Rate Estimation: A dictator does not need prior knowledge of the server's learning rate ($\eta$). They can numerically estimate it after a single round by sending a dummy update and observing the server's response.
Comparison of loss curves in a standard FL setting (left) versus a setting with one dictator client (right).

Collaborative and Betrayal Dynamics

The authors extend the threat model to groups of malicious actors. In a collaborative setting, multiple dictators work together to ensure the model converges on their shared interests. However, the paper reveals a fascinating game-theoretic twist: the Betrayal Mechanism. Even within a malicious coalition, a single client can secretly maintain a private shadow model to eventually "betray" their partners, erasing their contributions just as they did to the benign clients.

Loss curves for scenarios with two (left) and three (right) collaborative dictator clients.

Experimental Evaluation

The attacks were validated across Computer Vision and Natural Language Processing (NLP) benchmarks, including MNIST, CIFAR-10, and AG News (using a DistilBERT model). To highlight the "erasure" effect, the authors utilized a highly non-IID setup where each client held data for specific, non-overlapping classes.

Key Results

  • Total Erasure: In MNIST experiments, a single dictator reduced the global model's accuracy on all other clients' data to 0.00%, while maintaining ~99% accuracy on its own data.
  • NLP Dominance: On the AG News dataset, a dictator forced the DistilBERT model to achieve 100% accuracy on its own label while dropping performance on all other labels to 0%.
  • Defense Vulnerability: Common defenses like Gradient Norm Clipping proved largely ineffective against coalitions. If enough dictators work together, their combined updates still outweigh the benign clients despite clipping.

Discussion and Future Work

The dictatorship attack represents a unique challenge because it uses the server's own feedback loop to identify and subtract benign signals. Unlike standard poisoning, which adds noise, this is an "anti-signal" attack.

However, the current framework assumes full participation in every round (typical of "cross-silo" FL). Future research is needed to explore how random client sampling in "cross-device" FL or multiple local epochs (as in FedAvg) might mitigate these effects. This work opens the door for developing new robust aggregation rules specifically designed to detect the cancellation patterns characteristic of dictator updates.

Summary updated

Abstract

Federated learning (FL) has emerged as a promising paradigm for decentralized model training, enabling multiple clients to collaboratively learn a shared model without exchanging their local data. However, the decentralized nature of FL also introduces vulnerabilities, as malicious clients can compromise or manipulate the training process. In this work, we introduce dictator clients, a novel, well-defined, and analytically tractable class of malicious participants capable of entirely erasing the contributions of all other clients from the server model, while preserving their own. We propose concrete attack strategies that empower such clients and systematically analyze their effects on the learning process. Furthermore, we explore complex scenarios involving multiple dictator clients, including cases where they collaborate, act independently, or form an alliance in order to ultimately betray one another. For each of these settings, we provide a theoretical analysis of their impact on the global model's convergence. Our theoretical algorithms and findings about the complex scenarios including multiple dictator clients are further supported by empirical evaluations on both computer vision and natural language processing benchmarks.
BibTeX
@misc{alipour2025power,
  author = {Mohammadsajad Alipour and Mohammad Mohammadi Amiri},
  title = {{Power to the Clients: Federated Learning in a Dictatorship Setting}},
  year = {2025},
  eprint = {2510.22149},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.22149},
}
Figure