COMBAT: Alternated Training for Effective Clean-Label Backdoor Attacks

Authors: Tran Huynh, Dang Nguyen, Tung Pham, Anh Tran

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our backdoor attacks can reach near-perfect attack success rates and bypass all state-of-the-art backdoor defenses, as illustrated via comprehensive experiments on standard benchmark datasets. Our code is available at https://github.com/Vin AIResearch/COMBAT.
Researcher Affiliation Collaboration Tran Huynh1, Dang Nguyen1, 2, Tung Pham1, Anh Tran1 1Vin AI Research 2University of Maryland v.tranhn2@vinai.io, dangmn@umd.edu, v.tungph4@vinai.io, v.anhtt152@vinai.io
Pseudocode Yes Algorithm 1: COMBAT
Open Source Code Yes Our code is available at https://github.com/Vin AIResearch/COMBAT.
Open Datasets Yes We use three popular datasets, namely CIFAR-10 (Krizhevsky, Hinton et al. 2009), Image Net-10, and Celeb A (Liu et al. 2015), for our experiments.
Dataset Splits No The paper mentions training data and test data but does not explicitly provide details about a validation set split (e.g., percentages or counts).
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU model, CPU type, memory) used for running the experiments.
Software Dependencies No The paper mentions models and optimizers (e.g., U-Net, SGD optimizer) but does not provide specific version numbers for software libraries or dependencies (e.g., Python, PyTorch, TensorFlow).
Experiment Setup Yes Models are trained for 200 epochs using SGD optimizer. We use a batch size of 128 for CIFAR-10 and Celeb A and 32 for Image Net-10. The initial learning rate is set to 0.01 for CIFAR-10 and Celeb A, and 0.001 for Image Net-10, which is decreased tenfold at epoch 100 and 150. [...] We set λℓ2 and λd as 0.02 and 0.8, respectively. For the high-frequency removal tricks, we choose ratio r = 0.65 and use Gaussian blur filter with kernel size of 3 and standard deviation σ uniformly sampled from [0.1, 1].