AI Paper Atlas

Research map / Adversarial robustness and security / Backdoor and poisoning

CBD: A Certified Backdoor Detector Based on Local Dominant Probability

Zhen Xiang, Zidi Xiong, Bo Li

NeurIPS 2023 · Adversarial robustness and security › Backdoor and poisoning · approach: Trigger-based backdoor attacks · 4 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 LDP with adjustable conformal prediction for model-level detectionBackdoor detection · GTSRB, SVHN, CIFAR-10; BadNet, CB, BlendTPR (%): CBD0.2: 90, 95, 95; 95, 100, 90; 75, 100, 60Mixed
Use CBD rather than the strongest listed baseline for CB detectionBackdoor detection · SVHN CBTPR (%): 70 -> 100Improved
Use CBD rather than the strongest listed baseline for CB detectionBackdoor detection · CIFAR-10 CBTPR (%): 80 -> 100Improved
Use conformal CBD rather than supervised CBDsupBackdoor detection · GTSRB, SVHN, CIFAR-10 pooled table averageTPR average (%): 89.4 -> 88.9Hurt
Increase assumed outlier ratio from 0 to 0.2 (ablation)CBD · GTSRB BadNetTPR (%): 75 -> 90Improved
Increase assumed outlier ratio from 0.1 to 0.2 (ablation)CBD · GTSRB BadNetTPR (%): 90 -> 90No clear effect
Reduce shadow model count from 100 to 10 (ablation)CBD · GTSRB BadNetTPR (%): 90 -> 95No clear effect
Reduce shadow model count from 100 to 25 (ablation)CBD · GTSRB BadNetTPR (%): 90 -> 95No clear effect
Reduce shadow model count from 100 to 50 (ablation)CBD · GTSRB BadNetTPR (%): 90 -> 95No clear effect
Train shadow models on only 100 clean samples per class (ablation)CBD0.2 · CIFAR-10 BadNetTPR (%): 75 -> 35Hurt
With only 100 clean samples per class, increase assumed outlier ratio from 0.2 to 0.4 (ablation)CBD · CIFAR-10 BadNetTPR (%): 35 -> 60Improved
Use small, noise-resilient triggers rather than larger perturbation triggersCBD certification · Four benchmark datasetsCTPR (%): For ℓ2≤0.75: 98%, 84%, 98%, 40%; for 0.75<ℓ2≤1.5: 68%, 76%, 100%, 28%Mixed

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