| name | neurally-guided-adversarial-robustness |
| description | Dissociating spatial frequency reliance from adversarial robustness in neurally aligned DCNNs. Shows that adversarial robustness from neural alignment is NOT primarily driven by spatial frequency bias (LSF or human channel), but by deeper representational properties. Use when: analyzing neural alignment robustness, spatial frequency analysis of DCNNs, adversarial attack defense mechanisms, ventral visual stream modeling, or brain-inspired CNN robustness. Activation: neural alignment robustness, adversarial robustness DCNN, spatial frequency bias, human channel vision, ventral stream alignment, neurally guided network, brain-inspired adversarial defense, representational similarity analysis
|
Neurally-Guided Adversarial Robustness in DCNNs
Based on: Shao et al. (2026), arXiv:2605.04443
Problem
DCNNs aligned to human visual cortex activity show improved adversarial robustness,
but the mechanism is unclear. The leading hypothesis was that alignment shifts models
away from brittle high-frequency details toward low spatial frequencies (LSF).
Key Findings
- Neural alignment increases reliance on BOTH LSF and the mid-frequency "human
channel" (the narrow band critical for human object recognition)
- Dissociation discovered: Directly biasing models toward the human channel
(alone or with LSF) does NOT improve robustness — it even impairs it
- LSF bias produces only modest robustness gains despite much larger
spatial-frequency shifts than neurally aligned models achieve
- Spatial-frequency-biased models show little increase in similarity to human
neural representational geometry
- Conclusion: Altered spatial-frequency reliance is an emergent property of
learning more human-like representations, NOT the primary mechanism for robustness
Methodology
Spectral Bias Analysis
- Compute spatial frequency profiles of DCNN representations at each layer
- Compare LSF/mid-frequency content between standard and neurally aligned models
- Map frequency reliance to ventral visual stream alignment scores
Dissociation Experiment
- Steer DCNNs toward specific frequency bands during training
- Evaluate adversarial robustness (PGD, FGSM attacks) on frequency-biased models
- Compare robustness gains against neural alignment baseline
Representational Similarity
- Compute RSA between model representations and human fMRI/MEG data
- Test whether frequency bias recapitulates human neural geometry
- Analyze layer-by-layer representational changes
Implications
For Brain-AI Alignment
- Neural alignment confers robustness through representational structure,
not just frequency filtering
- Future work should examine representational properties beyond spatial-frequency
profiles: object-centricity, invariance structure, compositional coding
For Adversarial Defense Design
- Simple frequency filtering is insufficient for robustness
- Effective defenses must capture higher-order representational properties
of the human visual system
Related Skills
vlm-visual-cortex-alignment-robustness - VLM robustness through early visual cortex alignment
untrained-cnns-match-backprop-v1 - Untrained CNNs match backprop at V1
brain-inspired-snn-pattern-analysis - Brain-inspired computing patterns
dina-v1-population-activity-interpretation - V1 population activity analysis