Efficient Image-to-Image Diffusion Classifier for Adversarial Robustness

Authors: Hefei Mei, Minjing Dong, Chang Xu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct sufficient evaluations of the proposed classifier under various attacks on popular benchmarks. Extensive experiments show that our method achieves better adversarial robustness with fewer computational costs than DM-based and CNN-based methods. We perform extensive experiments to empirically demonstrate the superiority of IDC on various benchmarks.
Researcher Affiliation Academia Hefei Mei1, Minjing Dong*1, Chang Xu2 1City University of Hong Kong, China 2University of Sydney, Australia EMAIL, EMAIL, EMAIL
Pseudocode No The paper describes the methodology in detail using mathematical formulations and prose, but it does not include any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statements about providing source code, nor does it include any links to code repositories.
Open Datasets Yes We conduct experiments on the CIFAR-10, CIFAR-100 datasets (Krizhevsky, Hinton et al. 2009) and Tiny-Image Net (Deng et al. 2009).
Dataset Splits No The paper mentions evaluating "on the entire test dataset" for CIFAR-10, but it does not provide specific details on the training, validation, or test splits for any of the datasets used (CIFAR-10, CIFAR-100, Tiny-Image Net).
Hardware Specification Yes We train our classifier using the Adam optimizer with 256 batch size cross 4 Tesla V100-32GB GPUs, CUDA V10.2 in Py Torch V1.7.1 (Paszke et al. 2019).
Software Dependencies Yes We train our classifier using the Adam optimizer with 256 batch size cross 4 Tesla V100-32GB GPUs, CUDA V10.2 in Py Torch V1.7.1 (Paszke et al. 2019).
Experiment Setup Yes The diffusion timesteps are set to Ts = 4, and the learning rate is set to 0.0001. For the CIFAR-10 dataset, we train 400 epochs in total while we train 600 epochs for the CIFAR-100 dataset. The hyper-parameter α is set to 0.2. The model channels of the U-Net are set to cm = 64, and the number of Res Net blocks is set to n R = 1. For the CIFAR10 dataset, the upscale list of channels is set to u = [1, 4] while that for CIFAR-100 is set to u = [1, 4, 8].