Type

Feature

Description

This issue proposes the addition of a new aggregation and client weighting mechanism for Federated Learning, inspired by the paper:

FedRDF: A Robust and Dynamic Aggregation Function Against Poisoning Attacks in Federated Learning
Enrique Mármol Campos, Aurora González-Vidal, José L. Hernández-Ramos, and Antonio Skarmeta.
IEEE Transactions on Emerging Topics in Computing, vol. 13, no. 1, pp. 48–67, 2025.
DOI: 10.1109/TETC.2024.3474484

FedRDF introduces a robust and dynamic aggregation function that mitigates poisoning attacks in Federated Learning (FL).
Unlike traditional approaches such as Krum or Trimmed Mean, it computes pairwise similarities between client updates in the frequency domain using the Discrete Fourier Transform (DFT), and applies a dynamic rejection and weighting process to minimize the impact of malicious or inconsistent client updates — without relying on any manually defined threshold.

This PR enables the baseline proposed in
Issue #6039

Planned Implementation

1. Inherit from FedAvg

   • Reuse existing client sampling, evaluation, and parameter handling logic.

2. Override aggregate_fit() to implement robust, frequency-domain aggregation:

   • Compute pairwise spectral similarity among client updates using the Discrete Fourier Transform (DFT).
   • Weight each client adaptively based on its similarity to others (clients with anomalous spectra contribute less).
   • Perform inverse FFT to reconstruct the aggregated update in the spatial domain.
   • (Optional) Apply Kolmogorov–Smirnov test for skewness estimation when adaptive switching is enabled.

3. Key Parameters

   • threshold (float) – Optional parameter controlling adaptive switching between FedAvg and FedRDF:
     • 0.0: Always use DFT-based robust aggregation (fully dynamic and threshold-free).
     • > 0.0: Switch adaptively to DFT-based aggregation only when skewness exceeds this threshold.

4. Features

   • Threshold-free robust aggregation (dynamic weighting, no manual tuning).
   • Layer-wise DFT aggregation providing resistance to outlier and backdoor updates.
   • Fully compatible with existing Flower workflows and client APIs.

Additional Context

Reference

E. Mármol Campos, A. González-Vidal, J. L. Hernández-Ramos, and A. 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