AI Paper Atlas

Research map / Adversarial robustness and security / Backdoor and poisoning

Flatness-aware Sequential Learning Generates Resilient Backdoors

Hoang Pham, The-Anh Ta, Anh T Tran, Khoa D Doan

ECCV 2024 · Adversarial robustness and security › Backdoor and poisoning · approach: Training-time backdoor defense · 1 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
train backdoor with SBL (two-task sequential learning with SAM and CL) instead of conventional trainingBadNets, Blended, SIG, Dynamic · CIFAR-10 / GTSRB, fine-tuning defensesASR after defense: CIFAR-10 BadNets FT SGD-0.01 ASR 1.79 -> 100Improved
train backdoor with SBLBadNets on ImageNet-10 · ImageNet-10 after FT SGD-0.01ASR and clean accuracy: ASR 10.47 -> 76.54; CA 51.38 -> 85.19Improved
add continual learning regularization (Naive, EWC, Anchoring, AGEM) in step 1 without SAM (ablation)SBL on BadNets · CIFAR-10 FT SGD-0.01ASR: 1.79 -> 23.19 (Naive), 85.73 (Anchoring)Improved
add continual learning in step 1 without SAM (ablation)SBL on BadNets · CIFAR-10 FT-SAMASR: 2.5 -> 2.84 (Naive), 13.6 (AGEM)No clear effect
add SAM to conventional backdoor training (ablation)Conventional backdoor learning on BadNets · CIFAR-10 fine-tuning defensesASR: FT-SAM 2.5 -> 99.99; FT SGD-0.01 1.79 -> 3.14Mixed
add SAM in step 0 of SBL w. EWC (ablation)SBL w. EWC · CIFAR-10 BadNetsASR after FT-SAM: 6.13 -> 100Improved
use large step-1 learning rate 0.01 for naive CL (ablation)SBL w. Naive · CIFAR-10 BadNetsASR after FT SGD-0.01: 100 -> 6.34Hurt
use EWC at large step-1 learning rate 0.01 (ablation)SBL · CIFAR-10 BadNetsASR and CA: ASR 79.91 retained, step 1 CA 85.32Mixed
lower the poisoning rate to 1% (ablation)BadNets w. SBL · CIFAR-10ASR after FT-SAM: 99.80 (vs 4.01 for plain BadNets)No clear effect
apply SBL to other architectures (ablation)BadNets, Blended, SIG · CIFAR-10 ResNet-20 and VGG-16ASR after defenses: VGG-16 BadNets FT SGD-0.01 3.20 -> 100Improved

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