arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2004.05887cs.CL
收藏

频率引导的词替换检测文本对抗性示例

Frequency-Guided Word Substitutions for Detecting Textual Adversarial Examples

University College London(伦敦大学学院);Tilburg University(蒂尔堡大学)

Maximilian Mozes, Pontus Stenetorp, Bennett Kleinberg, Lewis D. Griffin

摘要:

最近的努力表明,神经文本处理模型容易受到对抗性示例的影响,但这些示例的性质却知之甚少。在这项工作中,我们展示了针对CNN、LSTM和基于Transformer的分类模型的对抗攻击,这些攻击通过被替换的单词及其对应的替换之间的频率差异来执行单词替换。基于这些发现,我们提出了频率引导的词替换(FGWS),一个简单的算法,利用敌对词替换的频率特性检测敌对的例子。FGWS通过准确检测SST-2和IMDb情感数据集上的对抗性示例实现了强大的性能,与基于RoberTa的分类模型相比,F1检测分数高达91.4%。我们比较我们的方法对最近提出的扰动歧视框架,并表明我们优于它高达13.0%F1。

英文摘要:

Recent efforts have shown that neural text processing models are vulnerable to adversarial examples, but the nature of these examples is poorly understood. In this work, we show that adversarial attacks against CNN, LSTM and Transformer-based classification models perform word substitutions that are identifiable through frequency differences between replaced words and their corresponding substitutions. Based on these findings, we propose frequency-guided word substitutions (FGWS), a simple algorithm exploiting the frequency properties of adversarial word substitutions for the detection of adversarial examples. FGWS achieves strong performance by accurately detecting adversarial examples on the SST-2 and IMDb sentiment datasets, with F1 detection scores of up to 91.4% against RoBERTa-based classification models. We compare our approach against a recently proposed perturbation discrimination framework and show that we outperform it by up to 13.0% F1.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑