AI Paper Atlas

Research map / Adversarial robustness and security / Backdoor and poisoning

Pre-activation Distributions Expose Backdoor Neurons

Runkai Zheng, Rongjun Tang, Jianze Li, Li Liu

NeurIPS 2022 · Adversarial robustness and security › Backdoor and poisoning · approach: Training-time backdoor defense · 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 BNP instead of ANP to defend BadNets (A2A)BNP · CIFAR-10 BadNets (A2A)ASR (%): 86.22 -> 0.72Improved
Use EP instead of NAD against CLAEP · CIFAR-10 CLAASR (%): 15.31 -> 0.91Improved
Use EP instead of ANP against CLAEP · CIFAR-10 CLAACC (%): 91.13 -> 94.42Improved
Prune backdoor neurons with EP against BadNetsResNet-18 · CIFAR-10 BadNets (A2O)ACC (%): 93.86 -> 93.88No clear effect
Use EP instead of ANP on WaNetEP · Tiny-ImageNet WaNetASR (%): 86.98 -> 0.02Improved
Use BNP instead of ANP on BadNets (A2O)BNP · Tiny-ImageNet BadNets (A2O)ACC (%): 53.85 -> 61.60Improved
Use BNP instead of Refool-case EP on Tiny-ImageNet RefoolEP · Tiny-ImageNet RefoolASR (%): 0.21 -> 17.07Mixed
Reduce benign samples for BNP from 500 to 10 (ablation)BNP · CIFAR-10 BadNets, CLA, BlendedACC/ASRNo clear effect
Increase number of benign samples for BNP (ablation)BNP · CIFAR-10 RefoolASRImproved
Use BNP instead of NAD for running timeBNP · 500 CIFAR-10 imagesrunning time (s): 22.08s -> 0.39sImproved
Use EP instead of FT for running timeEP · 500 CIFAR-10 imagesrunning time (s): 12.35s -> 10.69sImproved

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