SPDNet Riemannian PGD Adversarial Attacks

This repository accompanies the paper riemannian adversarial attacks on symmetric positive definite matrices. It focuses on Projected Gradient Descent (PGD) defined on the manifold of Symmetric Positive Definite (SPD) matrices using the affine-invariant Riemannian metric.

What is here

• Riemannian and Euclidean PGD/FGM attacks implemented in the notebook adversarial.ipynb (cells labelled "Attacks").
• Utility helpers to load models and data (utils.py) and to generate random SPD matrices (spd_random.py).

Requirements

• Python 3.10+
• PyTorch (CUDA optional), GeoOpt, MOABB, NumPy, Matplotlib...

Key files

• adversarial.ipynb: end-to-end evaluation and plotting of Riemannian PGD.
• utils.py: dataset/model loading and SPD checks.
• spd_random.py: SPD sampling utilities used for synthetic experiments.