Real-time multi-object tracking on embedded GPUs must process each frame within a strict per-frame budget. But availability attacks like SlowTrack can disrupt this by inflating the tracker's persistent state, causing the pipeline to miss deadlines and degrade system availability.
TrackShield, described in a new arXiv preprint, is a runtime defense built to counter such attacks. It is designed to keep high-performance camera-based perception pipelines within their real-time constraints even when under attack, according to the abstract.
The work targets resource-constrained embedded GPU platforms, where efficiency is critical. The abstract does not detail the defense mechanism, but it emphasizes that TrackShield is an efficient runtime solution against this class of availability threats.