Masked Adversarial Generation for Neural Machine Translation

TL;DR

This paper introduces MAG, a novel adversarial generator that enhances neural machine translation robustness by learning to produce meaningful attacks, outperforming existing methods in speed and effectiveness.

Contribution

The paper proposes the Masked Adversarial Generation (MAG) model, a new approach that learns to generate adversarial attacks during training, improving robustness of translation models.

Findings

MAG improves model robustness against adversarial attacks.

MAG is faster than existing adversarial attack methods.

MAG effectively perturbs translation models during training.

Abstract

Attacking Neural Machine Translation models is an inherently combinatorial task on discrete sequences, solved with approximate heuristics. Most methods use the gradient to attack the model on each sample independently. Instead of mechanically applying the gradient, could we learn to produce meaningful adversarial attacks ? In contrast to existing approaches, we learn to attack a model by training an adversarial generator based on a language model. We propose the Masked Adversarial Generation (MAG) model, that learns to perturb the translation model throughout the training process. The experiments show that it improves the robustness of machine translation models, while being faster than competing methods.