Adversarial Examples Can Be Effective Data Augmentation for Unsupervised Machine Learning
Authors: Chia-Yi Hsu, Pin-Yu Chen, Songtao Lu, Sijia Liu, Chia-Mu Yu6926-6934
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive experimental results show outstanding performance gains (up to 73.5% performance improvement) by retraining the model with UAEs. |
| Researcher Affiliation | Collaboration | Chia-Yi Hsu1, Pin-Yu Chen2, Songtao Lu2, Sijia Liu3, Chia-Mu Yu1 1National Yang Ming Chiao Tung University 2 IBM Research 3 Michigan State University |
| Pseudocode | Yes | Algorithm 1: Min Max Attack Algorithm |
| Open Source Code | Yes | Codes are available at https://github.com/IBM/UAE. |
| Open Datasets | Yes | We use the CIFAR-10 dataset and the same neural network as in Section 4.2 to provide qualitative and quantitative evaluations on the two per-sample MINE methods for image classification. ... We use the default implementation of the following four autoencoders to generate UAEs based on the training data samples of MNIST and SVHN for data augmentation... The six datasets and the resulting classification accuracy are reported in Table 6. We select M = 50 features for every dataset except for Mice Protein (we set M = 10) owing to its small data dimension. |
| Dataset Splits | No | The paper mentions training and testing data, but does not explicitly provide details about a distinct validation set split for reproducibility (e.g., percentages, counts, or methodology for creating a validation set). |
| Hardware Specification | No | The paper states, 'We provide a brief summary of the datasets and computing resource in Supp Mat 6.18.' However, the supplementary material (Supp Mat 6.18) is not included in the provided text, so specific hardware details are not available in the main paper. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as libraries or frameworks used in the experiments. |
| Experiment Setup | Yes | Min Max Algorithm Parameters We use consistent parameters by setting α = 0.01, β = 0.1, and T = 40 as the default values. |