Towards Robustness Certification Against Universal Perturbations

Authors: Yi Zeng, Zhouxing Shi, Ming Jin, Feiyang Kang, Lingjuan Lyu, Cho-Jui Hsieh, Ruoxi Jia

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Aside from an extensive evaluation of the proposed certification, we further show how the certification facilitates efficient comparison of robustness among different models or efficacy among different universal adversarial attack defenses and enables accurate detection of backdoor target classes. 5 EXPERIMENT
Researcher Affiliation Collaboration 1Virginia Tech, Blacksburg, VA 24061, USA 2University of California, Los Angeles, CA 90095, USA 3Sony AI, Tokyo, 108-0075, Japan
Pseudocode No The paper does not contain a clearly labeled "Pseudocode" or "Algorithm" block.
Open Source Code Yes 1https://github.com/ruoxi-jia-group/Universal_Pert_Cert
Open Datasets Yes For evaluating the certification, we consider two benchmark datasets, MNIST (Le Cun et al., 1998) and CIFAR-10 (Krizhevsky et al., 2009), widely adopted in existing works.
Dataset Splits Yes For evaluating the certification, we consider two benchmark datasets, MNIST (Le Cun et al., 1998) and CIFAR-10 (Krizhevsky et al., 2009), widely adopted in existing works.
Hardware Specification Yes We use one server equipped with a total of 8 RTX A6000 GPUs as the hardware platform.
Software Dependencies No Py Torch (Paszke et al., 2019) is adopted as the implementation framework. We use Gurobi (Bixby, 2007) to solve the MILP. While specific tools are named, their version numbers are not explicitly provided in the text.
Experiment Setup Yes We use Adadelta (Zeiler, 2012) as the optimizer with a learning rate set to 0.1 for all the model training process (including the adversarial training for the model updating step as well). For MNIST models, we train each model with 60 epochs. For CIFAR-10 models, we train each model with 500 epochs to ensure full convergence. For adversarial training adopted in the main text, the number of steps in PGD attacks is 7; step-size for PGD is set as ϵ/4.