Erosion Attack: Harnessing Corruption To Improve Adversarial Examples.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Cybersecurity

Background:

• Adversarial examples pose a significant threat to deep neural networks.
• Existing transferable adversarial attacks often fail against black-box defense models, masking the true threat.
• Current attacks are limited by data-dependency and network-overfitting.

Purpose of the Study:

• To propose a novel transferable attack capable of defeating diverse black-box defenses.
• To address the limitations of data-dependency and network-overfitting in adversarial attacks.
• To highlight the security vulnerabilities in current robust deep learning models.

Main Methods:

• Data Erosion: Identifies special augmentation data that behaves consistently across vanilla and defense models.
• Network Erosion: Extends a single surrogate model into a diverse ensemble to enhance attack transferability.
• Erosion Attack (EA): Integrates both Data and Network Erosion methods for superior attack transferability.

Main Results:

• Empirical evaluations demonstrate EA's superiority over existing transferable attacks.
• EA effectively defeats a wide range of black-box defenses.
• The study reveals significant security limitations in current robust models.

Conclusions:

• Erosion Attack (EA) presents a potent method for generating transferable adversarial examples.
• The proposed methods effectively mitigate data-dependency and network-overfitting issues.
• EA underscores the persistent and substantial threat adversarial examples pose to deep learning security.