Cloud providers increasingly offer FPGAs on demand, and multi-tenant FPGAs allow several users to share one device to improve utilization. That sharing creates a security risk: an adversary co-located on the same FPGA may be able to observe physical signals from other tenants' workloads.
The arXiv preprint describes ML-assisted power side-channel attacks on ML accelerators in shared FPGAs. The attack uses power side-channel measurements, aided by machine learning, to target accelerators running on the same device. The abstract frames this as a risk introduced by multi-tenant sharing, where the very feature that improves utilization also expands the attack surface.
The paper does not present a full exploit chain in the abstract, but it highlights that ML accelerators are a plausible target for co-located attackers. For cloud providers and FPGA users, the work is a reminder that sharing programmable hardware can create physical information leaks alongside the usual software security boundaries.