On Adversarial Examples for Character-Level Neural Machine Translation

Abstract

Evaluating on adversarial examples has become a standard procedure to measure robustness of deep learning models. Due to the difficulty of creating white-box adversarial examples for discrete text input, most analyses of the robustness of NLP models have been done through black-box adversarial examples. We investigate adversarial examples for character-level neural machine translation (NMT), and contrast black-box adversaries with a novel white-box adversary, which employs differentiable string-edit operations to rank adversarial changes. We propose two novel types of attacks which aim to remove or change a word in a translation, rather than simply break the NMT. We demonstrate that white-box adversarial examples are significantly stronger than their black-box counterparts in different attack scenarios, which show more serious vulnerabilities than previously known. In addition, after performing adversarial training, which takes only 3 times longer than regular training, we can improve the model's robustness significantly.

References in corpus (5)

• Ensemble Adversarial Training: Attacks and Defenses
• Character-Aware Neural Language Models
• Analysing How People Orient to and Spread Rumours in Social Media by Looking at Conversational Threads
• Generating Natural Adversarial Examples
• Understanding Adversarial Training: Increasing Local Stability of Neural Nets through Robust Optimization

Cited by in corpus (46)

• Adversarial Attacks on Deep Learning Models in Natural Language Processing: A Survey
• Natural Attack for Pre-trained Models of Code
• CharBERT: Character-aware Pre-trained Language Model
• Weight Poisoning Attacks on Pre-trained Models
• SemMT: A Semantic-based Testing Approach for Machine Translation Systems
• Robust Neural Machine Translation with Doubly Adversarial Inputs
• When and How to Fool Explainable Models (and Humans) with Adversarial Examples
• Towards a Robust Deep Neural Network in Texts: A Survey
• Combating Adversarial Misspellings with Robust Word Recognition
• Dirty Road Can Attack: Security of Deep Learning based Automated Lane Centering under Physical-World Attack
• A Targeted Attack on Black-Box Neural Machine Translation with Parallel Data Poisoning
• An Empirical Survey of Data Augmentation for Limited Data Learning in NLP
• Adversarial Reinforced Instruction Attacker for Robust Vision-Language Navigation
• Imitation Attacks and Defenses for Black-box Machine Translation Systems
• On Evaluation of Adversarial Perturbations for Sequence-to-Sequence Models
• A Reinforced Generation of Adversarial Examples for Neural Machine Translation
• Exploring the Robustness of NMT Systems to Nonsensical Inputs
• Findings of the First Shared Task on Machine Translation Robustness
• Robust Encodings: A Framework for Combating Adversarial Typos
• Detecting Overfitting via Adversarial Examples
• AdvAug: Robust Adversarial Augmentation for Neural Machine Translation
• Counterfactual Variable Control for Robust and Interpretable Question Answering
• Stress Test Evaluation of Transformer-based Models in Natural Language Understanding Tasks
• Visual Cues and Error Correction for Translation Robustness
• Defense of Word-level Adversarial Attacks via Random Substitution Encoding
• Can NMT Understand Me? Towards Perturbation-based Evaluation of NMT Models for Code Generation
• Contextualized word senses: from attention to compositionality
• Understanding Pre-Editing for Black-Box Neural Machine Translation
• From Hero to Zéroe: A Benchmark of Low-Level Adversarial Attacks
• Structure-Invariant Testing for Machine Translation
• Addressing the Vulnerability of NMT in Input Perturbations
• Masked Adversarial Generation for Neural Machine Translation
• Learning Neural Models for Natural Language Processing in the Face of Distributional Shift
• Improving Robustness of Malware Classifiers using Adversarial Strings Generated from Perturbed Latent Representations
• Mischief: A Simple Black-Box Attack Against Transformer Architectures
• Putting words into the system's mouth: A targeted attack on neural machine translation using monolingual data poisoning
• Adversarial Machine Learning in Text Analysis and Generation
• Testing Machine Translation via Referential Transparency
• On Robustness of Neural Semantic Parsers
• Adversarial Attacks Against Deep Learning Systems for ICD-9 Code Assignment
• Multi-granularity Textual Adversarial Attack with Behavior Cloning
• Contrasting Human- and Machine-Generated Word-Level Adversarial Examples for Text Classification
• Localized Uncertainty Attacks
• Evaluating Robustness to Input Perturbations for Neural Machine Translation
• Secoco: Self-Correcting Encoding for Neural Machine Translation
• Token Drop mechanism for Neural Machine Translation