Mind Control through Causal Inference: Predicting Clean Images from Poisoned Data

Authors: Mengxuan Hu, Zihan Guan, Yi Zeng, Junfeng Guo, Zhongliang Zhou, Jielu Zhang, Ruoxi Jia, Anil Vullikanti, Sheng Li

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our model can effectively and robustly recover the original true labels of backdoored images, without compromising clean accuracy. Our code can be found at https://github.com/xuanxuan03021/BKD BKD ICLR. Table 1 compares our method MCCI with other defense baselines across various backdoor attacks on two datasets.
Researcher Affiliation Collaboration 1University of Virginia 2Virginia Tech 3University of Maryland, College Park 4Merck & Co., Inc. 5University of Georgia
Pseudocode Yes Algorithm 1: Mind Control Through Causal Inference (MCCI)
Open Source Code Yes Our code can be found at https://github.com/xuanxuan03021/BKD BKD ICLR.
Open Datasets Yes Following (Guo et al., 2023a; Gao et al., 2019; Li et al., 2021a), we choose two widely-adopted datasets for evaluating the effectiveness of our proposed method: CIFAR10 (Krizhevsky, 2009), and Image Net-subset (Deng et al., 2009).
Dataset Splits Yes The details of the dataset are given in Table 5. Table 5: Statistical information about the Datasets Dataset Image Size # of Training samples # of Testing Samples # of Classes CIFAR-10 32 32 3 50,000 10,000 10 Image Net-Subset 224 224 3 9,469 3,925 10
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions specific tools and models like 'open-sourced backdoor learning toolbox (Li et al., 2023)', 'Vi T', 'CLIP', 'BLIP', 'Res Net-18', 'Efficient Net', but does not provide specific version numbers for any software dependencies (e.g., Python, PyTorch, TensorFlow, CUDA).
Experiment Setup Yes We use an initial learning rate of 0.1 that is decreased by a factor of 10 at epochs 30, 60, and 90, 100 epochs, a batch size of 128, and a weight decay of 1e-4 for training the defense model against all attack baselines.