Add Flower Baseline: FedRDF #6039
Description
Type
Feature
Description
Paper
Enrique Mármol Campos, Aurora Gonzalez-Vidal, José L. Hernández-Ramos, and Antonio Skarmeta
"FedRDF: A Robust and Dynamic Aggregation Function Against Poisoning Attacks in Federated Learning"
IEEE Transactions on Emerging Topics in Computing, vol. 13, no. 1, pp. 48–67, 2025.
DOI: 10.1109/TETC.2024.3474484
Motivation
FedRDF introduces a robust and dynamic aggregation function designed to mitigate poisoning attacks in federated learning.
Unlike static robust aggregators such as Krum or Trimmed Mean, FedRDF dynamically adjusts the contribution of each client based on similarity relationships derived from the clients’ model updates.
The function models client relationships using a Discrete Fourier Transform (DFT) over client updates to compute spectral similarities, applying a dynamic rejection and weighting process to minimize malicious influence without requiring an external threshold.
This baseline would strengthen Flower’s library of robust federated learning strategies, complementing ongoing work on client detection and defense mechanisms (see FedDefender and FLDetector).
Planned Implementation
Implementation Plan
-
New Strategy:
Implement a new aggregation strategy in:
flwr/server/strategy/fedrdf.py -
New Baseline Example:
Add a new baseline under:
baselines/fedrdf/
- Example dataset: MNIST or CIFAR-10
- Compare FedRDF vs FedAvg, TrimmedMean, and Krum
- Include example plots for accuracy and convergence
-
Integration:
The newFedRDFstrategy can leverage the client-results split mechanism introduced in
PR #5453,
enabling post-aggregation filtering and client-level scoring. -
Testing and Documentation:
- Minimal test in
baselines/tests/test_fedrdf.py - Add section to
baselines/README.md - Confirm linting (
make format), tests (make test), and documentation pass (make check)
Additional Context
Dependencies
Only standard libraries: torch, numpy, flwr, and optionally scipy for distance metrics.
Labels
New Baseline · Feature · Good First Issue