Sy-FAR: Symmetry-based Fair Adversarial Robustness

The artifact contains all components promised in the paper’s Open Science section, enabling full end-to-end reproduction of our results. Specifically, the repository includes:

Source Code (Training + Evaluation Pipelines)
Implements all methods described in the paper: baseline ERM and ROA adversarial training, FAAL reweighting, SpecNorm regularization, and the full Sy-FAR algorithm.
This matches the Open-Science commitment to release all training and fairness-aware methods used in the experiments.

Attack Implementations
Includes physically realizable attacks such as the eyeglass-frame attack and ROA, along with AutoAttack and randomized smoothing variants.
This fulfills the promise to release evaluation code for benign and physically realizable adversarial attacks.

Preprocessing Tools & Dataset Preparation Instructions
The repository includes scripts for face detection, alignment, and cropping (using FaceX-Zoo with a pinned commit), as well as instructions and templates for structuring the data into the exact layout used in the experiments. The artifact provides clear download instructions, preprocessing utilities, class lists, and folder structures. Together, these components ensure that users can faithfully reproduce all dataset preparation steps described in the paper.

Pretrained models We also include pre-trained models for all training schemes—standard training, ROA-based adversarial training, FAAL, SpecNorm, and our Sy-FAR method so users can directly evaluate clean and adversarial performance without retraining models from scratch.

Metrics & Visualization Suite
Tools for computing robust accuracy, fairness gaps, symmetry metrics, and target-class vulnerability, along with heatmap and exemplar visualizations.
These components correspond to the promised evaluation and reproduction tools for all reported metrics.

README and Documentation
A detailed README explains environment setup, dataset preparation, training commands, evaluation scripts, and reproduction steps.
This meets the commitment to provide step-by-step instructions for reproducing every result.

Together, these components fully match the Open-Science commitments stated in the accepted USENIX Security paper, ensuring transparency, reproducibility, and easy verification of all claims.