Attribute-Guided Adversarial Training for Robustness to Natural Perturbations
Authors: Tejas Gokhale, Rushil Anirudh, Bhavya Kailkhura, Jayaraman J. Thiagarajan, Chitta Baral, Yezhou Yang7574-7582
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the applicability of our approach on three types of naturally occurring perturbations objectrelated shifts, geometric transformations, and common image corruptions. Our approach enables deep neural networks to be robust against a wide range of naturally occurring perturbations. We demonstrate the usefulness of the proposed approach by showing the robustness gains of deep neural networks trained using our adversarial training on MNIST, CIFAR-10, and a new variant of the CLEVR dataset. 5 Experiments In this section, we introduce the three types of robustness specifications that we experiment with, along with details about the datasets, baselines, and metrics used for each. |
| Researcher Affiliation | Collaboration | 1 Arizona State University, 2 Lawrence Livermore National Laboratory {tgokhale, chitta, yz.yang}@asu.edu, {anirudh1, kailkhura1, jjayaram}@llnl.gov |
| Pseudocode | Yes | Algorithm 1 Attribute-Guided Adversarial Training |
| Open Source Code | No | The paper provides a link to a dataset (CLEVR-Singles), but it does not state that the source code for the described methodology is publicly available, nor does it provide a direct link to a code repository for their method. |
| Open Datasets | Yes | Dataset: To study the problem of such object-level shifts along semantic factors of an image in a controlled fashion, we create a new benchmark called CLEVR-Singles1 by modifying the data generation process from CLEVR (Johnson et al. 2017). Dataset: https://github.com/tejas-gokhale/CLEVR-Singles. Dataset: We address this problem in the digit classification setting, with the training images from MNIST (Le Cun et al. 1998)... Dataset The CIFAR10 dataset (Krizhevsky 2009) contains 50k training images belonging to 10 classes. |
| Dataset Splits | No | The paper describes training and testing procedures and some splits (e.g., train/test sets), but it does not explicitly detail a separate validation dataset split with specific percentages, counts, or methodology. |
| Hardware Specification | Yes | For each experiment, we train the Att GAN on the training dataset outlined in Table 1 to generate 128x128 images, with a learning rate of 2e 4 for 100 epochs on a single 16GB GPU. |
| Software Dependencies | No | The paper mentions deep learning models (e.g., ResNet-26) and techniques (e.g., Group Normalization, Spatial Transformer Networks) but does not provide specific version numbers for any software libraries or dependencies used (e.g., Python, TensorFlow, PyTorch versions). |
| Experiment Setup | Yes | All models are trained for 15 epochs including pre-training epochs Npre = 5, batch-size 64, and M = 15 update steps for adversarial augmentation. The number of augmented samples Taug is 30% of the original source data, and augmentation interval Naug is fixed at 2 epochs. For our model the coefficients in Equations 4, and 5 are: λ1 = 0.5, λ2 = 0.5, β = 0.25. The learning rates η, µ for the classifier and adversarial augmentation are both 5e-5. |