AI Paper Atlas

Research map / Adversarial robustness and security / Backdoor and poisoning

UMD: Unsupervised Model Detection for X2X Backdoor Attacks

Zhen Xiang, Zidi Xiong, Bo Li

ICML 2023 · Adversarial robustness and security › Backdoor and poisoning · approach: Trigger-based backdoor attacks · 3 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use UMD instead of K-ArmX2X backdoor detection · CIFAR-10 X2X attacks (Avg)MIA: 0.74 -> 0.91Improved
use UMD instead of K-ArmX2X backdoor detection · GTSRB X2X attacks (Avg)MIA: 0.87 -> 0.91Improved
use UMD instead of K-ArmX2X backdoor detection · Imagenette X2X attacks (Avg)MIA: 0.77 -> 0.85Improved
omit TR and apply MAD directly to all class-pair triggers (ablation)UMD · CIFAR-10 2to2, 5to5, 8to8, A2Ar (Pert trigger)Overall MIA: 0.93 -> 0.50Hurt
use TR without the proposed clustering technique (ablation)UMD · CIFAR-10 2to2, 5to5, 8to8, A2Ar (Pert trigger)Overall MIA: 0.93 -> 0.70Hurt
use the trigger reverse-engineering procedure with fewer than the default number of images (ablation)UMD · CIFAR-10MIAImproved
vary the targeted misclassification fraction below 1 (ablation)UMD trigger reverse-engineering · CIFAR-10MIANo clear effect
use UMD with WaNet or Blended triggersCIFAR-10 X2X backdoor detection · 2to2, 5to5, 8to8, A2ArMIA: WaNet: 0.80, 0.80, 1.0, 0.90; Blended: 0.90, 0.90, 0.90, 0.90Mixed
use UMD for adaptive backdoor detectionCIFAR-10 · ISSBA, CLA, N2N, O2OMIA: 1.0, 0.60, 1.0, 0.80Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy