Auth-Graph: GenAI-empowered attribute-masked backdoor for on-demand authorizable graph learning

Xiao Yang, Gaolei Li, Kai Zhou, Yuni Lai, Jianhua Li

Abstract: Highlights•We proposed the first authorizable graph learning methodology.•Our method is compatible with multi-user access environments with distinct tokens.•Our method efficiently grants authenticated legitimate access with high precision, while simultaneously rejecting the unauthorized.•Our method effectively denies malicious access even when the model is subject to adversarial piracy and attacks.