Learning how much to trust each client: Artificial Bee Colony (ABC) search for FedAvg aggregation weights under non-IID medical data.
In federated learning, hospitals train a shared model without sharing patient data. The standard aggregation rule, FedAvg, simply averages client models with fixed weights. When data is non-IID (each hospital sees a different mix of diseases), this fixed average pulls the global model in conflicting directions, and accuracy drops and fluctuates.
Instead of a fixed average, treat the aggregation weights as an optimization problem. After every communication round, the server runs ABC to search for the client weights w₁…w₅ that maximize accuracy on a small validation set, then aggregates with the normalized best weights:
θ_global = Σ (wᵢ / Σw) · θᵢ wᵢ ∈ [0.1, 1.0]
The lower bound of 0.1 keeps every hospital contributing, so no client is silently dropped.
| Dataset | PathMNIST (MedMNIST v2): colon pathology, 9 tissue classes, 28×28 RGB |
| Clients | 5 simulated hospitals |
| Non-IID split | Training set sorted by label and cut into 5 contiguous shards, so each client sees only a narrow subset of classes (strong label skew) |
| Scenarios | ① Centralized (upper bound, no privacy) · ② Standard FedAvg · ③ ABC-optimized FedAvg |
| Experiment A | 2-layer CNN, 10 rounds, 1 local epoch, ABC: 5 food sources, 15 evaluations per round |
| Experiment B | ResNet-18 (adapted to 28×28), 20 rounds, 3 local epochs, data augmentation, ABC: 30 evaluations per round, accuracy + macro-F1 |
| Method | Final test accuracy | Best round |
|---|---|---|
| Centralized (upper bound) | 78.7% | 80.6% |
| Standard FedAvg | 47.9% | 53.2% |
| ABC-FedAvg | 55.2% | 62.9% |
| Method | Final test accuracy | Final macro-F1 |
|---|---|---|
| Standard FedAvg | 49.3% | — |
| ABC-FedAvg | 53.7% | 0.502 |
ABC-FedAvg led FedAvg in accuracy in 17 of 20 rounds and stayed ahead on macro-F1 for most of training, which matters because accuracy alone can hide classes the model ignores.
- Learned aggregation weights consistently closed part of the gap between FedAvg and the centralized upper bound.
- The learned weights change from round to round, e.g.
[0.22, 0.04, 0.23, 0.25, 0.27]in the last round of Experiment A, showing that the server learns to down-weight clients whose updates hurt the global model. - A large gap to centralized training remains, which reflects how hard this label-skewed split is.
git clone https://github.com/yusufkorkmazyigit/fedLearningwithABC.git
cd fedLearningwithABC
pip install -r requirements.txt
jupyter notebook abc_optimized_fedAvg.ipynbPathMNIST downloads automatically on first run. A GPU is recommended, especially for the ResNet-18 experiment (tested on Google Colab).
This is an exploratory study, and the following points are known limitations:
- Validation data: the 300 samples ABC uses to score weight combinations are taken from the test split. A cleaner protocol would carve the validation set out of training data (or from a server-side holdout) so the final test set stays untouched.
- Single run: results come from one seed. Federated training on non-IID data is noisy, so mean ± std over several seeds is the next step.
- ABC variant: the notebook uses a simplified ABC loop with a small evaluation budget for speed. The canonical implementation is available in abc-optimizer-lib.
- Baselines: comparing against FedProx, FedNova and loss-based weighting would place the method in the literature.
- Cost: each ABC evaluation runs inference on the validation set, which adds server-side computation per round.
- abc-optimizer-lib: canonical ABC with a scikit-learn wrapper
- McMahan et al. (2017). Communication-Efficient Learning of Deep Networks from Decentralized Data (FedAvg)
- Karaboğa, D. (2005). An Idea Based on Honey Bee Swarm for Numerical Optimization, Erciyes University
- Yang et al. (2023). MedMNIST v2: A Large-Scale Lightweight Benchmark for 2D and 3D Biomedical Image Classification, Scientific Data
Developed by Yusuf Korkmazyiğit, Computer Engineering, Erciyes University.