Version	1.0.0
Main Libraries	Flower PyTorch Matplotlib
Testing Datasets	torchvision.datasets torchtext.datasets torchaudio.datasets
Testing Models	torchvision.models (+120)

FLiP

This repository extends AP4Fed (Architectural Patterns for Federated Learning), a Flower-based reference implementation equipped with an adaptation subsystem that enables and disables architectural patterns and tunes their parameters across FL rounds. The implementation builds on the ideas introduced in FLiP [1], where architectural patterns are toggled at runtime (round-by-round) to better match evolving system conditions and performance goals.

Authors

• Luciano Baresi (Politecnico di Milano)
• Ivan Compagnucci (Gran Sasso Science Institute)
• Livia Lestingi (Politecnico di Milano)
• Catia Trubiani (Gran Sasso Science Institute)

FLiP in a Nutshell

FLiP is a Federated Learning framework built on top of Flower, an open-source Python library designed to simplify the development of Federated Learning systems. FLiP enhances this foundation with an adaptation layer that monitors each training round and selects which architectural patterns to apply, enabling or disabling them and tuning their parameters to match current system conditions and performance goals.

Experiments for [1]

_JSS__Adaptive_FL_Patterns__ECS…

• Detailed Dockerized implementation and adaptation docs: Docker/README.md

Results analysis (Docker/results)

The Docker/results folder contains the experimental results. Below is a brief description of scripts processing results to address three research questions.

RQ1 — FLiP Effectiveness (Docker/results/rq1.py)

• What it does: Performs RQ1 comparative analysis across architectural patterns and activation strategies (e.g., selector, heterogeneous_data_handler (hdh), compressor). It aggregates metrics from replicated runs, filters scenarios (e.g., N_high vs N_low mixes), generates plots, and runs non-parametric statistical tests.

• Methods: Mann–Whitney U test for pairwise comparisons and Vargha–Delaney A effect size. Includes outlier filtering and per-round vs cumulative metric views.

• Outputs: Prints statistical summaries to console and, if enabled, saves figures under Docker/results/plots/rq1/.

• Quickstart:

  # From the repository root
  python3 Docker/results/rq1.py

RQ2 — Offline Overhead (Docker/results/rq2.ipynb)

• What it does: Evaluates the quality of predictor-based components used in adaptation by fitting/assessing simple predictive models per pattern (selector, hdh, compressor).

• Outputs: Saves regression quality plots (RMSE and R-squared) under Docker/results/plots/rq2/ (e.g., rq2_rmse_selector.pdf, rq2_rsquared_selector.pdf, analogous for hdh and compressor).

• Quickstart:

  # Option A: Execute the notebook headlessly and save an executed copy
  jupyter nbconvert --to notebook --execute Docker/results/rq2.ipynb \
    --output Docker/results/rq2.executed.ipynb
  
  # Option B: Open interactively
  jupyter notebook Docker/results/rq2.ipynb

RQ3 — Online Overhead (Docker/results/rq3.ipynb)

• What it does: Analyzes time-related behavior across configurations/patterns, e.g., round duration and communication/training time components, and compares distributions across scenarios.

• Outputs: Saves boxplots under Docker/results/plots/rq3/ (e.g., rq3_time_boxplot.pdf).

• Quickstart:

  # Option A: Execute the notebook headlessly and save an executed copy
  jupyter nbconvert --to notebook --execute Docker/results/rq3.ipynb \
    --output Docker/results/rq3.executed.ipynb
  
  # Option B: Open interactively
  jupyter notebook Docker/results/rq3.ipynb

If you are looking to run the experiments and produce the input data for these analyses, see Docker/README.md.

References

[1] Baresi, L., Compagnucci, I., Lestingi, L., & Trubiani, C. (2026). Adaptive Toggling of Architectural Patterns for Federated Learning. Accepted for the International Conference on Software Engineering for Adaptive and Self-Managing Systems (SEAMS) 2026. Rio de Janeiro, Brazil, 13–14 April 2026.

@inproceedings{BaresiSEAMSFlIP,
  author    = {Baresi, Luciano and Compagnucci, Ivan and Lestingi, Livia and Trubiani, Catia},
  title     = {{Adaptive Toggling of Architectural Patterns for Federated Learning}},
  booktitle = {International Conference on Software Engineering for Adaptive and Self-Managing Systems (SEAMS)},
  year      = {2026},
  address   = {Rio de Janeiro, Brazil},
  date      = {13--14 April 2026}
}