Research2min read

Denial of Deadline: Network-Driven Accuracy Collapse in Distributed Inference Pipelines

This paper identifies a new attack vector scenario called 'Denial of Deadline', where targeted workload bursts push the latency limit for the slow, high-accuracy inference path beyond its deadline. This causes high-quality predictions to be discarded and drastically reduces overall system utility.

High relevanceAI SecurityModel SecurityAdversarial Machine Learning

Research problem

Distributed inference architectures often combine a fast path for real-time requirements with a slow, compute-intensive path on cloud hardware to maximize accuracy. The coordination layer between these paths has been little examined against coordinated workload attacks.

Methodology

The authors conduct an experimental evaluation in a simulated two-tier edge-cloud pipeline for autonomous driving to analyze the effect of 'Yo-Yo' bursts on latency and accuracy.

Key findings

  • Targeted burst attacks can increase p99 latencies at benign traffic from 92ms up to 2 seconds, nearly eliminating the benefit of the slow path.
  • Accuracy can be degraded by an average of 7.0 HOTA points through such attacks, with certain rare classes (e.g., stop signs) losing nearly half their precision.
  • The attack requires no access to model weights or victim data and targets solely temporal synchronization.

Practical impact

For security architects, this means that resource isolation and strict latency guarantees in the coordination layers are critical. The results motivate research into defenses against workload contention to ensure the 'slow path' maintains its accuracy benefits under load.

Limitations

  • The study is based on simulations; real conditions in heterogeneous edge-cloud environments may vary.
  • No specific countermeasures or patches are proposed, only research directions are identified.

Details

Original title
Denial of Deadline: Network-Driven Accuracy Collapse in Distributed Inference Pipelines
Authors
Jhonatan Tavori, Gur-Eyal Sela, Ion Stoica, Gil Zussman
Institutions
University of California, Berkeley, Technion - Israel Institute of Technology, Microsoft Research
Publication platform
arXiv
Methodology type
experimental-evaluation
Paper
Paper

Why it matters

This paper is relevant for AI security practitioners because it describes a new attack model that does not rely on classical adversarial machine learning against the model operators themselves. Instead, the attacker exploits weaknesses in infrastructure coordination of distributed systems to degrade service through accuracy loss (Accuracy Collapse via DoS). For teams implementing routing, merging, or scheduling mechanisms for distributed inference, this represents a systemic risk vector often overlooked.

Sources