Open Science Artifact: Leveraging Runtime Explanation Similarity for Architectural Patterns in Federated Learning

This package contains the scripts and results of the experiments conducted for the paper "Leveraging Runtime Explanation Similarity for Architectural Patterns in Federated Learning", which is currently under revision. The package contains the following file: README.md: Instructions to reproduce the entire study pipeline. To reproduce the pipeline described in this study and illustrated in Figure 3 of the paper, please follow the instructions in the README.md. (Please do not change the file locations, as each script depends on specific file paths.) figures/: Folder for saving output plots. models/: Folder containing the original .pt model files (for both clients and server) and the .csv file with evaluation metrics from the standard experiment considered in the paper. Note that this folder already contains the original output files of the experiments described in the paper, enabling exact reproduction of the results. 1-4 Federated Learning Execution/: Folder containing the SSIM4Fed Federated Learning Benchmarking Platform used to run FL experiments. The tool already integrates the SSIM-based selection strategy implemented in Stage 4 of the pipeline. To reproduce the Federated Learning process and generate new .pt files, please refer to the README-Tool.md file included in this folder. 2-3 Visual Explanation and Model Similarity Computation/: Folder containing scripts to generate visual explanations using Grad-CAM and to compute model similarity using the SSIM metric. paper_figures.ipynb: A Jupyter Notebook file used to preprocess and visualize the experimental results based on the output generated by the scripts in the 2-3 Visual Explanation and Model Similarity Computation/ folder. Note: Although the replication package is relatively large, this size is necessary to include the full set of input images and intermediate artifacts required to faithfully reproduce all experiments and visualizations of the explanations. Because these data must be stored and processed locally, hardware with limited storage, memory, or compute capabilities may struggle to execute the entire pipeline or may incur significantly longer processing times.