Presentation of “Breaking GNN-based NIDS via Hard-Label Black-Box Flow Injection Attacks” at the PTTC
Date:
Presentation of the paper “Breaking GNN-based NIDS via Hard-Label Black-Box Flow Injection Attacks” at the PTTC (Programme de Transfer du Campus Cyber) on September 24, 2026. The talk was presented for an mixed audience of researchers and industry professionals.
While Graph Neural Network (GNN)-based Network Intrusion Detection Systems (NIDS) excel by encoding relational traffic struc- tures, this structural dependency introduces novel adversarial vulner- abilities. Existing graph attacks remain impractical, often demanding white-box access, prohibitive query budgets, or exploit-breaking payload manipulations. To overcome these limitations, we propose a practical, hard-label black-box attack that evades detection strictly by injecting valid supplementary traffic flows, perfectly preserving the exploit’s func- tional integrity. By formulating this combinatorial challenge as a weakly submodular maximization problem, we leverage a local surrogate GNN and an influence-guided greedy search to efficiently identify highly impactful injections. We theoretically prove this approach yields a strict polynomial-time approximation guarantee. Empirically, we demonstrate the proposed flow injection attack can bypass state-of-the-art GNN-based NIDS with minimal query budgets