Disttack: Graph Adversarial Attacks Toward Distributed GNN Training

本文介绍了Disttack，这是首个针对分布式图神经网络（GNN）训练的对抗攻击框架，利用了分布式系统中频繁的梯度更新特性。具体而言，Disttack通过向单个计算节点注入对抗性攻击来破坏分布式GNN训练。被攻击的子图被精确扰动以在反向传播中引起异常的梯度上升，破坏计算节点之间的梯度同步，从而导致训练GNN的性能显著下降。我们在四个大型真实世界的图上评估了Disttack，攻击了五个广泛采用的GNN。实验结果表明，与现有的对抗攻击方法相比，Disttack将模型准确度下降的幅度增加了2.75倍，并平均加速了17.33倍，同时保持了不可察觉性。

Disttack: Adversarial Attacks for Distributed Graph Neural Networks

Disttack is the first framework of adversarial attacks for distributed GNN training that leverages the characteristics of frequent gradient updates in a distributed system. It corrupts distributed GNN training by injecting adversarial attacks into one single computing node, disrupting gradient synchronization between computing nodes and leading to a significant performance decline of the trained GNN.

Disttack is evaluated on four large real-world graphs by attacking five widely adopted GNNs. Experimental results demonstrate that Disttack amplifies the model accuracy degradation by 2.75x and achieves speedup by 17.33x on average while maintaining unnoticeability.

Related work in this field includes adversarial attacks on GNNs, distributed machine learning, and privacy-preserving distributed learning.