Security Research Primitives and Attack Simulations in Flower

Contributions
1

Hi everyone,

We’re exploring the idea of introducing security research primitives into Flower to support research on robustness and adversarial behavior in federated learning. This would include standardizing and expanding support for common attack types, as well as enabling systematic evaluation of defense strategies

In many FL papers, malicious clients are used to simulate real-world threats. Attacks such as backdoors, scaling attacks, label flipping, Byzantine strategies, and more are essential for testing robustness.

We’d love your input! Here’s how you can help:

Ideas and Design

Suggest how attacks and defenses should be structured or modularized in Flower. Feel free to also share papers, datasets, or tools that simulate or defend against attacks in FL.

Code and Prototypes

Help develop reusable attack components or extend existing examples with malicious behaviors. We will help you throughout Pull Request/Contribution process.

Discussion

  • What types of attacks and defenses should we prioritize?
  • How can we evaluate detection, mitigation, and impact of attacks in a standardized way?

Looking forward to discuss this topic!

2

I’m interested in contributing to the Flower framework, especially at this topic. But before I submit a large PR with multiple simulated attacks, I’d like to better understand the scope and goals of this contribution topic.

For example: Which attacks should be simulated? Should the simulation produce machine-readable output, defined by a JSON Schema, so it can be easily integrated into MLOps workflows? Which security scoring/benchmarking system should be implemented? Do you have specific papers in mind that describe attacks which should be implemented as simulations in the Flower framework?

Would you recommend opening a new issue with an outline of my proposed ideas so we can discuss the architecture there? Or is there a specific existing issue on this topic where I could start contributing?