New HQA-VLAttack Framework Targets Vision-Language AI Models with High-Quality Adversarial Examples
A new research paper presents HQA-VLAttack, a framework designed to create high-quality adversarial examples for vision-language pre-trained models. It tackles black-box attacks, which only utilize model predictions, by considering both text and image alterations. Existing approaches often depend on resource-heavy iterative methods or concentrate solely on positive image-text pairs, which hampers their effectiveness. HQA-VLAttack functions in two phases: generating text perturbations through counter-fitting word vectors and executing image attacks. This study addresses the early stage of adversarial attack research, aiming to improve query efficiency and similarity reduction. The framework aims to bolster the robustness testing of multimodal AI systems. This research was published on arXiv under identifier 2604.16499v1, contributing to the analysis of AI model vulnerabilities.
Key facts
- HQA-VLAttack is a framework for generating adversarial examples on vision-language pre-trained models
- The approach addresses black-box attacks where only predicted results are accessible
- Existing methods either use complex iterative cross-search strategies requiring many queries
- Current techniques often ignore reducing similarity of negative image-text pairs
- The framework consists of text and image attack stages
- Text perturbation generation leverages counter-fitting word vectors
- Research on this specific adversarial attack problem is described as being in its infancy
- The paper was published on arXiv with identifier 2604.16499v1
Entities
Institutions
- arXiv