DAFA: Distance-Aware Fair Adversarial Training

Authors: Hyungyu Lee, Saehyung Lee, Hyemi Jang, Junsung Park, Ho Bae, Sungroh Yoon

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results across various datasets demonstrate that our method not only maintains average robust accuracy but also significantly improves the worst robust accuracy, indicating a marked improvement in robust fairness compared to existing methods. We conduct both theoretical and empirical analyses of robust fairness, taking into account inter-class similarity. Experiments across multiple datasets validate that our approach effectively boosts the worst robust accuracy.
Researcher Affiliation Academia Hyungyu Lee1, Saehyung Lee1, Hyemi Jang1, Junsung Park1, Ho Bae2,*, and Sungroh Yoon1,3,* 1Electrical and Computer Engineering, Seoul National University 2Department of Cyber Security, Ewha Womans University 3Interdisciplinary Program in Artificial Intelligence, Seoul National University
Pseudocode Yes A THE ALGORITHMS OF DAFA Algorithm 1 Training procedure of DAFA ... Algorithm 2 DAFAcomp
Open Source Code Yes Our code is available at https://github.com/rucy74/DAFA.
Open Datasets Yes We conducted experiments on CIFAR-10, CIFAR-100 (Krizhevsky et al., 2009), and STL-10 (Coates et al., 2011)
Dataset Splits No The paper mentions 50,000 training images and 10,000 test images for CIFAR-10/100, and 5,000 training images and 8,000 test images for STL-10. It does not explicitly define a separate validation set split or describe a validation procedure for its own experiments, apart from mentioning prior work that uses training and validation performances.
Hardware Specification Yes Our experiments were carried out on a single RTX 8000 GPU equipped with CUDA11.6 and Cu DNN7.6.5.
Software Dependencies Yes Our experiments were carried out on a single RTX 8000 GPU equipped with CUDA11.6 and Cu DNN7.6.5.
Experiment Setup Yes We set the learning rates to 0.1, implementing a decay at the 100th and 105th epochs out of a total of 110 epochs, using a decay factor of 0.1 as recommended by Pang et al. (2021). For optimization, we utilized stochastic gradient descent with a weight decay factor of 5e-4 and momentum set to 0.9. The upper bounds for adversarial perturbation were determined at 0.031 (ϵ = 8). The step size for generating adversarial examples for each model was set to one-fourth of the ℓ -bound of the respective model, over a span of 10 steps. For our method, the warm-up epoch was set to τ = 70 and the hyperparameter λ was set to λ = 1 for CIFAR10 and λ = 1.5 for CIFAR-100 and STL-10 due to the notably low performance of hard classes.