Generalization Bound and New Algorithm for Clean-Label Backdoor Attack

Authors: Lijia Yu, Shuang Liu, Yibo Miao, Xiao-Shan Gao, Lijun Zhang

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we empirically validate the proposed backdoor attack on benchmark datasets CIFAR10, CIFAR100 (Krizhevsky et al., 2009), SVHN and Tiny Image Net(Le & Yang, 2015), and against popular defenses. We also conduct ablation experiments to verify our main Theorems 4.1 and 4.5.
Researcher Affiliation Academia 1Institute of Software, Chinese Academy of Sciences, Beijing 100190, China 2State Key Laboratory of Computer Science 3Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing 100190, China 4University of Chinese Academy of Sciences, Beijing 100049, China 5Kaiyuan International Mathematical Sciences Institute.
Pseudocode Yes Algorithm 1 provides detailed steps for creating the trigger, where is element-wise product.
Open Source Code Yes Code is in https://github.com/hong-xian/backdoor-attack.git.
Open Datasets Yes We empirically validate the proposed backdoor attack on benchmark datasets CIFAR10, CIFAR100 (Krizhevsky et al., 2009), SVHN and Tiny Image Net(Le & Yang, 2015), and against popular defenses.
Dataset Splits No The paper mentions training on a “training set” and evaluating on a “test set” but does not explicitly specify a validation set split (e.g., percentages or counts for training, validation, and test sets).
Hardware Specification Yes We do our experiments on Pytorch and GPU NVIDIA Ge Force RTX 3090.
Software Dependencies No The paper mentions “Pytorch” but does not provide a specific version number for Pytorch or any other software library or solver used in the experiments.
Experiment Setup Yes When we train victim network, we use SGD, we have 150 epochs in the training, the learning rate is 0.01, and reduce to 80% at 40-th,80-th, 120-th epochs, use weight decay 10^-4, momentum 0.9, each data in the training set will flip or randomly crop before inputting network in the training.