Adversarial Attack Type I: Cheat Classifiers by Significant Changes.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) are highly successful but vulnerable to adversarial attacks.
• Existing attacks often involve small perturbations, increasing Type II errors (false negatives).

Purpose of the Study:

• To introduce and evaluate a novel Type I adversarial attack.
• To demonstrate that this attack causes significant data alterations while maintaining misclassification.
• To differentiate Type I and Type II adversarial attacks.

Main Methods:

• A supervised variation autoencoder was designed to generate Type I adversarial examples.
• Latent variables were updated using gradient information to attack classifiers.
• Type I attacks on latent spaces were explored using pre-trained generative models.

Main Results:

• The proposed method effectively generates Type I adversarial examples on large-scale image datasets.
• Generated examples often bypass detectors designed for Type II attacks.
• Attack strategies effective against one type are not necessarily effective against the other.

Conclusions:

• Type I and Type II adversarial attacks are fundamentally different, stemming from distinct underlying reasons.
• The proposed Type I attack poses a significant threat to DNNs by causing substantial data changes.
• Current defense mechanisms may need to be re-evaluated for their effectiveness against diverse adversarial attack types.