Adversarial Distributional Training for Robust Deep Learning

Authors: Yinpeng Dong, Zhijie Deng, Tianyu Pang, Jun Zhu, Hang Su

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on several benchmarks validate the effectiveness of ADT compared with the state-of-the-art AT methods.
Researcher Affiliation Collaboration Yinpeng Dong , Zhijie Deng , Tianyu Pang, Jun Zhu, Hang Su Dept. of Comp. Sci. & Tech., Institute for AI, BNRist Center Tsinghua-Bosch Joint ML Center, THBI Lab, Tsinghua University, Beijing, 100084 China
Pseudocode Yes We provide the general algorithm for ADT in Alg. 1. Algorithm 1 The general algorithm for ADT
Open Source Code Yes Code is available at https://github.com/dongyp13/Adversarial-Distributional-Training.
Open Datasets Yes We perform experiments on the CIFAR10 [34], CIFAR-100 [34], and SVHN [44] datasets.
Dataset Splits No The paper mentions using test sets for evaluation but does not explicitly detail the train/validation/test dataset splits, such as percentages or sample counts.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running its experiments.
Software Dependencies No The paper mentions using Adam for optimization but does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes Training Details: We adopt the cross-entropy loss as L in our objective (5). We set λ = 0.01 for the entropy term, and leave the study of the effects of λ in Sec. 5.3. For ADTEXP, we adopt Adam [32] for optimizing φi with the learning rate 0.3, the optimization steps T = 7, and the number of MC samples in each step k = 5.