Weijia Shao, William Eriksson Kling, Simon Gunnarsson, Muoi Tran · ACM SIGSAC Conference on Computer and Communications Security (CCS) 2026 · 2026
DOI: 10.1145/3830454.3846436
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
The host network has emerged as a major source of performance loss for distributed AI workloads in the data center. We show that an unprivileged tenant can weaponize this fabric, issuing only ordinary memory traffic to congest the shared network paths and slow the training steps of a co-located victim workload. Our tested attack inflates Transformer training by 2.3× and GNN training by 3.7×, without any elevated privilege, exploit, or escape. These results establish the host network as an integral part of the attack surface of multi-tenant AI and call for a deeper study of its mitigation.
No comments yet — start the discussion below.