Dissociating spatial frequency reliance from adversarial robustness advantages in neurally guided deep convolutional...

Area of Science:

• Computer Vision
• Neuroscience
• Artificial Intelligence

Background:

• Deep convolutional neural networks (DCNNs) excel at visual tasks but are vulnerable to adversarial attacks.
• Aligning DCNN representations with human visual cortex activity improves adversarial robustness.
• The exact mechanisms behind this neural alignment advantage remain unclear.

Purpose of the Study:

• Investigate whether low spatial frequencies (LSF) or the "human channel" drives adversarial robustness in neurally aligned DCNNs.
• Determine if spectral bias is the primary mechanism for robustness or an emergent property.

Main Methods:

• Aligned DCNNs to higher-order human ventral visual stream regions.
• Directly steered DCNNs towards LSF and the human channel.
• Assessed adversarial robustness and similarity to human neural representational geometry.

Main Results:

• Neural alignment increased reliance on both LSF and the human channel.
• Directly biasing models towards the human channel impaired robustness.
• LSF bias yielded modest robustness gains but little similarity to human neural geometry.

Conclusions:

• Altered spatial-frequency reliance is likely an emergent property of learning human-like representations.
• Neural alignment confers robustness through mechanisms beyond spatial-frequency bias.
• Future research should explore representational properties beyond spatial-frequency profiles.