ADVERSARIAL ATTACK ON NEURAL MACHINE TRANSLATION MODELS AS DISCRETE OPTIMIZATION

• Main Author: Kuwanto, Garry
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/65291
• Institution: Institut Teknologi Bandung

Machine Translation models are commonly trained only on perfect grammar corpora. Training on perfect grammar corpora might predispose the model to be biased against minorities from non-standard linguistics backgrounds. Adversarial attack is one way to protect machine learning models from this bias. In this final project, We perform an adversarial attack by means of perturbation of words using the different morphological forms. This perturbation crafts adversarial examples with non-perfect grammar for English sentences. Because of how choosing the morphological forms is seen as finding an optimal choice from a discrete search space, we can see this as a discrete optimization problem. Here, we focus on exploring three different discrete optimization method, Genetic Algorithm, Simulated Annealing, and a modified version of a reinforcement learning algorithm, the REINFORCE algorithm. Comparison with the greedy algorithm as a commonly used method, our results show little difference between sentence crafted from these methods. In terms of efficiency, Simulated Annealing is a significantly faster algorithm, three times faster than the slowest algorithm (Genetic Algorithm) and faster than the second best algorithm (REINFORCE) by a factor of two.