[论文解读] Data Augmentation Using GANs

Fabio Henrique Kiyoiti dos Santos Tanaka, Claus Aranha|arXiv (Cornell University)|Apr 19, 2019

Digital Media Forensic Detection参考文献 12被引用 143

一句话总结

本文使用 Generative Adversarial Networks 生成用于训练分类器的合成数值数据并平衡不均衡数据集；结果在某些情况下与原始数据的准确率/召回率媲美，且相对于基线有改进，尽管在高度不平衡的任务中 SMOTE/ADASYN 可能优于基于 GAN 的过采样。

ABSTRACT

In this paper we propose the use of Generative Adversarial Networks (GAN) to generate artificial training data for machine learning tasks. The generation of artificial training data can be extremely useful in situations such as imbalanced data sets, performing a role similar to SMOTE or ADASYN. It is also useful when the data contains sensitive information, and it is desirable to avoid using the original data set as much as possible (example: medical data). We test our proposal on benchmark data sets using different network architectures, and show that a Decision Tree (DT) classifier trained using the training data generated by the GAN reached the same, (and surprisingly sometimes better), accuracy and recall than a DT trained on the original data set.

研究动机与目标

• Motivate data augmentation to address imbalanced datasets and privacy concerns.
• Evaluate GAN-generated synthetic data as training data for classifiers.
• Assess GAN-based oversampling against SMOTE and ADASYN.
• Identify GAN architectures that yield effective synthetic data for numerical tabular data.

提出的方法

• Utilize a GAN to generate synthetic numerical data that mirrors original data distributions.
• Train a Decision Tree classifier on synthetic data and compare to original-data training.
• Experiment with six GAN configurations by varying network depth and width.
• Balance datasets by oversampling the minority class using GAN-generated data and compare to SMOTE/ADASYN.
• Preprocess data with min-max scaling to [0,1] before GAN training.
• Assess similarity via mean Euclidean distance between synthetic and original data points.

实验结果

研究问题

• RQ1 Can GAN-generated synthetic data train a classifier with comparable or better performance than training on original data?
• RQ2 Can GANs be effectively used to balance imbalanced datasets compared to SMOTE and ADASYN?
• RQ3 What GAN architectures (depth/width) yield the best performance for numeric, non-image datasets?
• RQ4 Do synthetic data improve privacy by reducing direct leakage of original data attributes?

主要发现

• A 256/512-layer GAN architecture generally yielded the best overall outcomes across datasets, with statistical significance for accuracy (p<0.05).
• In some cases, classifiers trained on GAN-synthetic data achieved accuracy and precision close to or better than those trained on original data.
• GAN-based oversampling improved results relative to the original imbalanced data but did not consistently beat SMOTE or ADASYN on highly imbalanced tasks (credit card fraud), especially in recall-sensitive settings.
• Training on fully synthetic data can sometimes preserve class distributions and attributes without explicit class-separation during GAN training.
• Euclidean distance analyses suggest synthetic data can be sufficiently distinct from the original to offer privacy advantages, especially in the cancer dataset.