AI Paper Atlas

Research map / Adversarial robustness and security / Backdoor and poisoning

Augmented Neural Fine-tuning for Efficient Backdoor Purification

Nazmul Karim, Abdullah Al Arafat, Umar Khalid, Zhishan Guo, Nazanin Rahnavard

ECCV 2024 · Adversarial robustness and security › Backdoor and poisoning · approach: Training-time backdoor defense · 6 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
add mask regularizer to NFT (ablation)NFT · CIFAR-10 LIRA, 10 samplesACC: 74.58 -> 82.72Improved
add MixUp augmentation to NFT (ablation)NFT · CIFAR-10 LIRA, 10 samplesASR: 12.23 -> 5.18Improved
add MixUp augmentation to NFT (ablation)NFT · CIFAR-10 LIRA, 10 samplesACC: 81.05 -> 82.72Hurt
add mask scheduling function mu(l) (ablation)NFT · CIFAR-10 LIRA, 100 samplesASR: 4.33 -> 2.04Improved
fine-tune neural masks with MixUp instead of adversarial trigger search baselinesNFT vs AWM · CIFAR-10 LIRA, 10 samplesASR: 91.45 -> 5.18Improved
use NFT instead of ANP/AWM adversarial search for defensebackdoor purification · CIFAR-10 runtimeruntime (sec): 118.1 -> 28.4Improved
apply NFT on ImageNet scaleNFT · ImageNetaverage ASR drop / ACC drop: 94.03 / 2.84Improved
choose alpha 0.8, beta 0.50 for exponential mask schedule (ablation)NFT · CIFAR-10 LIRAASR/ACC: ASR 1.53, ACC 90.57Improved
NFT in multi-label detection settingNFT · VOC12ASR: RNP 13.8 vs NFT 14.2Mixed

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