Neural alignment's adversarial robustness not solely due to low spatial frequency bias
A new study investigates why deep convolutional neural networks (DCNNs) aligned with human visual cortex activity show improved adversarial robustness. Previous work suggested this robustness comes from a bias toward low spatial frequencies (LSF), but recent findings indicate human object recognition relies on a narrow mid-frequency "human channel" partially preserved in prior LSF-focused studies. The researchers examine whether spectral bias toward LSF or the human channel primarily drives adversarial robustness in neurally aligned DCNNs. They first show that DCNNs aligned to higher-order regions exhibit specific spectral properties. The study challenges the assumption that LSF bias alone explains robustness gains, pointing to the human channel's role. The preprint is available on arXiv under identifier 2605.04443.
Key facts
- DCNNs aligned with human visual cortex activity show improved adversarial robustness.
- One hypothesis attributes this robustness to bias toward low spatial frequencies.
- Human object recognition critically depends on a narrow mid-frequency 'human channel'.
- Prior LSF-focused studies partially preserved this human channel.
- The study investigates whether LSF or human channel bias drives robustness.
- DCNNs aligned to higher-order regions show specific spectral properties.
- The preprint is on arXiv with identifier 2605.04443.
- The research challenges the LSF bias explanation for adversarial robustness.
Entities
Institutions
- arXiv