Humans can decipher adversarial images.

Area of Science:

• Cognitive Science
• Computer Vision
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) achieve human-level image classification.
• CNNs serve as models for human vision, but differ in susceptibility to adversarial examples.
• Adversarial examples are patterns or perturbations that cause AI misclassification, challenging their reliability.

Purpose of the Study:

• To investigate whether human and machine judgments on adversarial images diverge.
• To determine if human intuition can predict machine classification of adversarial images.

Main Methods:

• Conducted 8 experiments using 5 diverse adversarial image datasets.
• Presented human subjects with adversarial images and assessed their classification predictions.
• Compared human judgments against machine classifications and relevant foils.

Main Results:

• Human subjects consistently and accurately predicted the machine's preferred label for adversarial images.
• This held true even for images described as unrecognizable to humans.
• Human and machine classifications of adversarial images demonstrated a robust relationship.

Conclusions:

• Human intuition is a surprisingly reliable indicator of machine (mis)classification.
• The findings have implications for understanding both human cognition and AI behavior.
• This suggests potential commonalities between human and artificial visual processing systems.