HyGloadAttack: Hard-label black-box textual adversarial attacks via hybrid optimization.

Area of Science:

• Natural Language Processing
• Machine Learning Security
• Adversarial Machine Learning

Background:

• Hard-label black-box textual adversarial attacks are challenging due to text's discrete nature and limited model access.
• Existing methods like exchange-based and gradient-based attacks suffer from local optima and high query requirements.
• Generating high-quality adversarial examples with semantic similarity and low perturbation under limited queries remains an open problem.

Purpose of the Study:

• To propose a novel framework, HyGloadAttack, for generating high-quality adversarial text examples.
• To address the limitations of existing methods in terms of performance and efficiency for textual adversarial attacks.
• To enhance semantic similarity and minimize perturbations in adversarial examples under constrained query conditions.

Main Methods:

• HyGloadAttack employs a perturbation matrix in word embedding space for global initialization.
• It selects synonyms to maximize similarity while preserving adversarial properties.
• A gradient-based quick search accelerates the optimization process.

Main Results:

• HyGloadAttack demonstrates significant superiority over state-of-the-art baseline methods.
• Experiments were conducted on five datasets for text classification and natural language inference.
• The method was also validated on two real-world APIs.

Conclusions:

• HyGloadAttack effectively generates high-quality adversarial text examples.
• The proposed framework overcomes limitations of previous methods in efficiency and perturbation control.
• This work advances the field of hard-label black-box textual adversarial attacks.