ETV-Attack: Efficient text-driven visual-variable adversarial attacks on visual question answering with pre-trained language models

Key Points

• Adversarial attacks can significantly disrupt visual question answering systems, affecting their reliability.
• Key evidence includes over 70% accuracy drop in targeted scenarios, indicating severe vulnerabilities.
• This analysis employs text-driven methods to systematically assess visual variable impact on question answering models.
• Findings highlight the need for robust defenses in AI systems to mitigate adversarial vulnerabilities.