Annealing Self-Distillation Rectification Improves Adversarial Training

Authors: Yu-Yu Wu, Hung-Jui Wang, Shang-Tse Chen

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the efficacy of ADR through extensive experiments and strong performances across datasets. In this section, we compare the proposed ADR to PGD-AT and TRADES in Table 1. We further investigate the efficacy of ADR in conjunction with model-weight-space smoothing techniques Weight Average (WA) (Izmailov et al., 2018; Gowal et al., 2020) and Adversarial Weight Perturbation (AWP) (Wu et al., 2020) with Res Net18 (He et al., 2016a) in Table 2. Experiments are conducted on well-established benchmark datasets, including CIFAR-10, CIFAR-100 (Krizhevsky et al., 2009), and Tiny Image Net-200 (Le & Yang, 2015; Deng et al., 2009).
Researcher Affiliation Academia Yu-Yu Wu, Hung-Jui Wang, Shang-Tse Chen National Taiwan University {r10922018,r10922061,stchen}@csie.ntu.edu.tw
Pseudocode Yes Algorithm 1 Annealing Self-Distillation Rectification (ADR)
Open Source Code Yes Furthermore, the source code can be found in the supplementary materials to ensure the reproducibility of this project.
Open Datasets Yes Experiments are conducted on well-established benchmark datasets, including CIFAR-10, CIFAR-100 (Krizhevsky et al., 2009), and Tiny Image Net-200 (Le & Yang, 2015; Deng et al., 2009).
Dataset Splits Yes During training, we evaluate the model with PGD-10 and select the model that has the highest robust accuracy on the validation set with early stopping (Rice et al., 2020).
Hardware Specification Yes The experiment is reported by running each algorithm on a single NVIDIA RTX A6000 GPU with batch size 128.
Software Dependencies No The paper mentions software components like "SGD optimizer" and uses models like "Res Net-18" and "WRN-34-10", but it does not specify any version numbers for programming languages (e.g., Python), deep learning frameworks (e.g., PyTorch, TensorFlow), or other key libraries.
Experiment Setup Yes We perform adversarial training with perturbation budget ϵ = 8/255 under l -norm in all experiments. In training, we use the 10-step PGD adversary with step size α = 2/255. We adopt β = 6 for TRADES as outlined in the original paper. The models are trained using the SGD optimizer with Nesterov momentum of 0.9, weight decay 0.0005, and a batch size of 128. The initial learning rate is set to 0.1 and divided by 10 at 50% and 75% of the total training epochs. Simple data augmentations include 32 32 random crop with 4-pixel padding and random horizontal flip (Rice et al., 2020; Gowal et al., 2020; Pang et al., 2021) are applied in all experiments. Following Wu et al. (2020); Gowal et al. (2020), we choose radius 0.005 for AWP and decay rate γ = 0.995 for WA. For CIFAR-10/100, we use 200 total training epochs, λ follows cosine scheduling from 0.7 to 0.95, and τ is annealed with cosine decreasing from 2.5 to 2 on CIFAR-10 and 1.5 to 1 on CIFAR100, respectively. As for Tiny Image Net-200, we crop the image size to 64 64 and use 80 training epochs. We adjust λ from 0.5 to 0.9 and τ from 2 to 1.5 on this dataset.