New Perspective-Invariant Attack Enhances Adversarial Example Transferability

Researchers have developed a new method called Perspective-Invariant Attack (PIA) to enhance the transferability of adversarial examples in deep neural networks. Unlike previous methods that used limited local transformations, PIA employs a multi-degree-of-freedom vertex sampling strategy to cover a hierarchy of perspective transformations, from simple translation to complex projective mapping. This approach generates geometrically diverse variations, reducing adversarial perturbation overfitting to surrogate models and improving cross-model transferability. A further extension, PIA-Mix, combines perspective transformations with auxiliary methods for even greater effectiveness, outperforming existing state-of-the-art attacks across various network architectures and multimodal large language models. AI

IMPACT This research could lead to more robust defenses against adversarial attacks on AI models, particularly those integrating multimodal data.

RANK_REASON Academic paper detailing a new method for adversarial attacks on deep neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• adversarial example
• deep neural network
• Multimodal Large Language Models
• Perspective-Invariant Attack
• PIA-Mix

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →