Hard label adversarial attack with high query efficiency against NLP models

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computer Security

Background:

• Black-box adversarial attacks effectively reveal vulnerabilities in natural language processing (NLP) models.
• Current attack methods are inefficient due to high query counts, limiting practical application.

Purpose of the Study:

• To propose QEAttack, a query-efficient hard-label attack method for NLP models.
• To generate persuasive and semantically equivalent adversarial texts with minimal queries.

Main Methods:

• Leveraging a genetic algorithm for adversarial text generation.
• Employing a dual-gradient fusion strategy to reduce query counts during crossover.
• Utilizing locality-sensitive hashing for sentence-level semantic clustering in mutation.

Main Results:

• QEAttack achieves high attack success rates across diverse NLP models and datasets.
• Demonstrates significantly reduced query counts compared to existing methods.
• Maintains or improves the imperceptibility and quality of adversarial texts.

Conclusions:

• QEAttack offers an efficient and effective solution for hard-label black-box adversarial attacks.
• The proposed strategies successfully address the query inefficiency of current adversarial text generation techniques.
• QEAttack enhances the practical applicability of adversarial attacks in NLP model robustness evaluation.