A universal adversarial policy for text classifiers.

Area of Science:

• Natural Language Processing
• Machine Learning
• Adversarial Learning

Background:

• Universal adversarial perturbations significantly impact adversarial learning theory and practice.
• Existing universal attacks in text often create unnatural inputs by altering all texts identically.
• The vision domain has seen success with universal perturbations, but text presents unique challenges due to linguistic nuances.

Purpose of the Study:

• To introduce a novel universal adversarial setup for text: a universal adversarial policy.
• To develop an attack that maintains text validity and naturalness, unlike previous universal text attacks.
• To demonstrate the existence and effectiveness of universal adversarial patterns in the text domain.

Main Methods:

• Learning a single search policy over a set of semantics-preserving text alterations.
• Utilizing synonym replacements as a perturbation strategy, known for naturalness in non-universal attacks.
• Employing a reinforcement learning approach as a strong baseline for policy learning.

Main Results:

• The universal adversarial policy successfully generates adversarial examples on new, unseen texts efficiently.
• The approach demonstrates generalization capabilities, effective even with a small training dataset (as few as 500 texts).
• The learned policy produces valid and natural-sounding adversarial texts, addressing limitations of prior universal text attacks.

Conclusions:

• Universal adversarial patterns are confirmed to exist in the text domain.
• The proposed universal adversarial policy offers a practical and effective method for generating adversarial text.
• This research advances adversarial learning in NLP by providing a more natural and versatile universal attack strategy.