This repository contains a proof-of-concept implementation demonstrating adversarial vulnerability in complex-valued Graph Attention Network (GAT) beamformers for multi-user MISO (MU-MISO) wireless networks.
The architecture under study is based on Reference [57] in:
Lu, Y., Zhang, S., Liu, C., Zhang, R., Ai, B., Niyato, D., Ni, W., Wang, X., & Jamalipour, A. (2025). Agentic Graph Neural Networks for Wireless Communications and Networking Towards Edge General Intelligence: A Survey. arXiv:2508.08620
Finding 1: Adversarial Vulnerability A complex FGSM attack on CSI input achieves 70.3% sum rate degradation at perturbation strength ε=0.05 — smaller than typical channel estimation error in real deployments.
| Condition | Sum Rate |
|---|---|
| Clean baseline | 23.95 bits/s/Hz |
| ε = 0.05 (attack) | 7.12 bits/s/Hz |
| Degradation | 70.3% |
Finding 2: Phase Blindness The trained model fails to learn global phase equivariance despite SINR being phase-invariant by physical law.
| Metric | Value |
|---|---|
| Equivariance score (ideal) | 0.0 |
| Equivariance score (ours) | 1.2901 |
Standard GAT with LeakyReLU activations cannot be made phase-equivariant through loss regularization alone. The architecture requires equivariant-by-design layers — identified as an open problem in Section VII-B of the survey.
- Architecture: Complex-valued GAT with residual connections
- Dataset: DeepMIMO O1 scenario, 28 GHz, 250,000 user locations
- BS antennas: 64-element ULA
- Users per graph: K = 8
- Training: Unsupervised Lagrangian loss, 2000 steps
- Attack: Complex FGSM on CSI input
phase-blind-gat/ complex_gat_fgsm_deepmimo.ipynb ← main notebook figures/ attack_curve.png ← Figure 1: robustness curve convergence_curves.png ← Figure 2: training convergence pe_training_curves.png ← Figure 3: PE regularization README.md
Preliminary implementation. Results obtained on a single run without fixed random seed. Reproducible version in progress.
If referencing this work, please also cite the survey:
bibtex @article{lu2025agentic, title={Agentic Graph Neural Networks for Wireless Communications and Networking Towards Edge General Intelligence: A Survey}, author={Lu, Yang and Zhang, Shengli and Liu, Chang and Zhang, Ruichen and Ai, Bo and Niyato, Dusit and Ni, Wei and Wang, Xianbin and Jamalipour, Abbas}, journal={arXiv preprint arXiv:2508.08620}, year={2025} }