graph-foundation-models

This paper identifies and attacks the alignment layer of graph foundation models, showing it is a distinct attack surface vulnerable to perturbations at low budgets, especially for spectral tokenizers, and proposes detection-based defenses.

Proposes FedGAMMA, a federated multimodal graph foundation learning framework that aligns multimodal attributes and graph topology via two-stage pre-training and prompt-based fine-tuning, achieving significant gains on multiple datasets.