UMD: Unsupervised Model Detection for X2X Backdoor Attacks

Authors: Zhen Xiang, Zidi Xiong, Bo Li

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct comprehensive evaluations on CIFAR-10, GTSRB, and Imagenette dataset, and show that our unsupervised UMD outperforms SOTA detectors (even with supervision) by 17%, 4%, and 8%, respectively, in terms of the detection accuracy against diverse X2X attacks.
Researcher Affiliation Academia Zhen Xiang 1 Zidi Xiong 1 Bo Li 1 1University of Illinois at Urbana-Champaign.
Pseudocode Yes Algorithm 1 UMD against X2X backdoor attacks
Open Source Code Yes The code related to this work can be found at: https://github.com/polaris-73/MT-Detection
Open Datasets Yes Dataset: We consider three benchmark image datasets, CIFAR-10 (Krizhevsky, 2012), GTSRB (Stallkamp et al., 2012), and Imagenette (Deng et al., 2009)
Dataset Splits Yes In our experiments, we follow the standard train-test split for each dataset (see Apdx. C.1 for details).
Hardware Specification Yes Empirically, each model inference on CIFAR-10, GTSRB, and Imagenette requires around 0.3h, 2.5h, and 4.3h, respectively, as measured on a single RTX 2080 Ti card.
Software Dependencies No The paper mentions 'Res Net-18' for model architecture and 'Adam optimizer' for training, but does not provide specific version numbers for software libraries or dependencies like Python or PyTorch.
Experiment Setup Yes Detailed training configurations are shown in Apdx C.3.