GateDrain: Availability Attacks and Admission-Side Defense for Confidence-Gated Edge-Cloud Inference
GateDrain lowers model confidence to force edge-cloud offload, inflating shared-queue latency for benign users.
GateDrain is an availability attack on confidence-gated edge-cloud inference: bounded input perturbations lower calibrated confidence so requests that would be answered locally are offloaded into a shared cloud queue without increasing the application request rate. Evaluations on the public EdgeBoost artifact cover white-box, transfer, decision-only, universal, and multi-gate attacks, and adaptive attacks defeat the tested training-free preprocessing defenses. The authors evaluate Bounded Escalation, combining per-source admission budgets, protected capacity, and non-preemptive trusted-class priority, plus an optional global bucket. Authenticated clients gain latency isolation, while tighter containment can reject legitimate unauthenticated offloads and reduce expected accuracy through edge fallback.
- Perturbations lower confidence and redirect local requests to the cloud.
- Application request rate stays fixed while cloud demand rises.
- Adaptive attacks bypass evaluated training-free preprocessing defenses.
- Bounded Escalation uses admission budgets and trusted-class priority.
- Tighter containment can reject legitimate unauthenticated offloads.
- Products
- EdgeBoost
From the original article
1 min read at arxiv.orgConfidence-gated edge--cloud inference accepts confident local predictions and offloads uncertain inputs to a stronger cloud model. We show that this routing decision creates an availability attack surface.
By Zonghua Gu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.