Improving Adversarial Robustness via Promoting Ensemble Diversity
Authors: Tianyu Pang, Kun Xu, Chao Du, Ning Chen, Jun Zhu
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results on various datasets verify that our method can improve adversarial robustness while maintaining state-of-the-art accuracy on normal examples. |
| Researcher Affiliation | Academia | 1Department of Computer Science and Technology, Institute for AI, BNRist Center, THBI Lab, Tsinghua-Fuzhou Institute for Data Technology, Tsinghua University, Beijing, China. Correspondence to: Tianyu Pang <pty17@mails.tsinghua.edu.cn>, Ning Chen <ningchen@tsinghua.edu.cn>, Jun Zhu <dcszj@tsinghua.edu.cn>. |
| Pseudocode | Yes | The training procedure with the ADP regularizer is called the ADP training method, as described in Algorithm 1. Algorithm 1 The ADP training procedure |
| Open Source Code | Yes | The code is available at https://github.com/P2333/Adaptive-Diversity-Promoting. |
| Open Datasets | Yes | We choose three widely studied datasets MNIST (Le Cun et al., 1998), CIFAR10 and CIFAR-100 (Krizhevsky & Hinton, 2009). |
| Dataset Splits | No | Each dataset has 50,000 training images and 10,000 test images. |
| Hardware Specification | Yes | The weight initialization is different for each network. We apply Adam optimizer (Kingma & Ba, 2014) with an initial learning rate of 0.001. Following similar setting as in He et al. (2016), we separately run the training process for 40 epochs on MNIST, 180 epochs on CIFAR-10 and CIFAR-100 with the mini-batch size of 64 on a Tesla P100 GPU worker. |
| Software Dependencies | No | The paper mentions using 'Adam optimizer' but does not specify any software libraries (e.g., PyTorch, TensorFlow) or their version numbers, which are necessary for full reproducibility of the software environment. |
| Experiment Setup | Yes | We apply Adam optimizer (Kingma & Ba, 2014) with an initial learning rate of 0.001. Following similar setting as in He et al. (2016), we separately run the training process for 40 epochs on MNIST, 180 epochs on CIFAR-10 and CIFAR-100 with the mini-batch size of 64 on a Tesla P100 GPU worker. |