Federated Learning, Agentic AI

InteFL

Knowledge Management Framework for Federated Learning

A configuration and execution framework for federated learning simulations built on Flower. Implements novel PID controller-inspired and Trust-based aggregation strategies alongside classical robust methods (Krum, Multi-Krum, Bulyan, RFA, Trimmed-Mean), adversarial attack scheduling, federated LLM fine-tuning with LoRA, and a wide range of medical imaging datasets. Published in IEEE Intelligent Systems, 2026.

Built with Python, Flower, PyTorch, LoRA

Information Flyer ↗ (PDF, 3 MB, opens in a new tab)

GitHub ↗IEEE DOI ↗

Experiments

Updated

2,502 experiments since March 2026. 2,472 ran to completion, using 166 hours of compute.

Latest run

Started
Dataset
FEMNIST (IID)
Aggregation strategy
PID-MADE
Experiments per weekBusiest week: Apr 6, 2026, 585 experiments
Show the weekly counts as a table
Experiments per week
Week ofExperiments
Mar 2, 20266
Mar 9, 202612
Mar 16, 202675
Mar 23, 2026258
Mar 30, 202695
Apr 6, 2026585
Apr 13, 2026333
Apr 20, 2026186
Apr 27, 2026183
May 4, 202670
May 11, 20260
May 18, 20260
May 25, 20260
Jun 1, 20260
Jun 8, 20260
Jun 15, 20260
Jun 22, 202623
Jun 29, 2026125
Jul 6, 20260
Jul 13, 202656
Jul 20, 20260
Jul 27, 20261
Aug 3, 20260
Aug 10, 20260
Aug 17, 20260
Aug 24, 20260
Aug 31, 202639
Sep 7, 20267
Sep 14, 20266
Sep 21, 202632
Sep 28, 2026304
Oct 5, 202654
Oct 12, 202621
Oct 19, 202615
Oct 26, 20260
Nov 2, 202616

Attacks tested

Data poisoning

Malicious clients corrupt their training data.

  • Label flipping1,061
  • Target flipping664
  • Gaussian noise187

Model poisoning

Malicious clients tamper with the model updates they send.

  • Gradient scaling260
  • A Little Is Enough156

Other attacks

  • Threshold evasion6

No attack

Every client is honest, for comparison.

  • Clean baseline287

Datasets

Image classification

  • FEMNISTIID: 875, non-IID: 3461,221
  • CIFAR-100IID: 57, α 0.1: 64, α 0.5: 59, α 2.0: 60, α 5.0: 1241
  • CIFAR-10IID: 84, non-IID: 10, α 0.1: 19, α 0.5: 24, α 2.0: 14151

Sensor time series

  • SMORES dissolved oxygen467
  • Air Quality (UCI)394

Medical imaging

  • PneumoniaMNIST16
  • BloodMNIST9
  • OCTMNIST2
  • Lung CT (NLST)1

α is the Dirichlet concentration used to split a dataset across clients. A smaller α gives each client a more uneven mix of classes.

Chart and breakdowns from InteFL's experiment database as of Nov 6, 2026, 11:00 AM EST. Get the data as JSON, or the latest run.

Contributors