Hierarchical binding in convolutional neural networks: Making adversarial attacks geometrically challenging.

Area of Science:

• Neuroscience
• Computer Vision
• Machine Learning

Background:

• Current deep learning models for machine vision are vulnerable to adversarial examples, which are subtle, often imperceptible image perturbations.
• This vulnerability may stem from the lack of explicit feature binding representations, unlike in the primate brain.
• Adversarial examples are theorized to arise from 'off-manifold' perturbations in image data.

Purpose of the Study:

• To develop a novel deep-learning architecture that improves machine vision robustness.
• To incorporate hierarchical feature binding mechanisms, inspired by primate visual neuroscience, into convolutional neural networks.
• To investigate the relationship between feature binding, representational geometry, and robustness against adversarial attacks.

Main Methods:

• Proposed a novel deep-learning architecture inspired by theoretical neuroscience on primate visual feature binding.
• Implemented hierarchical feature binding representations within convolutional neural networks.
• Empirically evaluated the architecture's robustness against various L0, L2, and Linf adversarial attacks, especially in black-box settings.

Main Results:

• The novel architecture demonstrated enhanced robustness against a range of adversarial attacks, particularly in the black-box setting.
• Analysis confirmed that the observed robustness was not due to gradient masking (false robustness).
• Representational geometry analysis revealed a positive correlation between hierarchical binding, expanded data manifolds, and robustness.

Conclusions:

• The proposed architecture, by approximating hierarchical feature binding, offers improved robustness in machine vision systems.
• Robustness appears to emerge from preserving both general low-level and abstract features.
• The findings suggest a potential link between hierarchical binding, human sensitivity to adversarial examples, and enhanced model robustness.