Denial of Deadline: Network-Driven Accuracy Collapse in Distributed Inference Pipelines
Burst-shaped traffic can make distributed inference miss its deadline and quietly lose the accuracy it was built to add.
The paper describes “accuracy collapse” in systems that pair a timely fast path with a more accurate remote slow path. In their autonomous-driving tracking simulation, about 4,000 shaped burst requests pushed benign p99 latency from 92ms to 2s, causing slow-path results to be discarded. Tracking quality fell by 7.0 HOTA points on average, with targeted intervals showing losses from 2.0 to 18.7 points. The attack does not require model weights or victim data. ArXiv · AI/CL/LG's note
The paper describes “accuracy collapse” in systems that pair a timely fast path with a more accurate remote slow path. In their autonomous-driving tracking simulation, about 4,000 shaped burst requests pushed benign p99 latency from 92ms to 2s, causing slow-path results to be discarded. Tracking quality fell by 7.0 HOTA points on average, with targeted intervals showing losses from 2.0 to 18.7 points. The attack does not require model weights or victim data. ArXiv · AI/CL/LG's note
score 4