Recognizing Object by Components With Human Prior Knowledge Enhances Adversarial Robustness of Deep Neural Networks

Area of Science:

• Computer Vision
• Cognitive Psychology
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial attacks, limiting their reliability in object recognition.
• Current DNNs lack part-based inductive bias, unlike human recognition, contributing to weak adversarial robustness.
• Existing defense methods against adversarial attacks are often adaptively evaded.

Purpose of the Study:

• To propose a novel object recognition model, ROCK (Recognizing Object by Components with human prior Knowledge), inspired by human cognitive psychology.
• To enhance the adversarial robustness of object recognition systems by incorporating human-like part-based processing.

Main Methods:

• ROCK employs a two-stage process: first, segmenting objects into parts, and second, scoring these segmentations using predefined human prior knowledge.
• The model mimics the human recognition-by-components theory, involving object decomposition and a human-like decision process.
• ROCK's performance is evaluated against classical recognition models under various adversarial attack settings.

Main Results:

• ROCK demonstrates superior robustness compared to classical recognition models when subjected to diverse adversarial attacks.
• The proposed part-based approach shows significant potential in improving the resilience of object recognition systems.
• The findings suggest that incorporating human prior knowledge enhances model defense capabilities.

Conclusions:

• The ROCK model offers a promising alternative to current DNN-based object recognition systems, particularly in adversarial scenarios.
• Part-based models, previously overlooked, hold significant potential for developing more robust AI systems.
• Future research should explore the integration of cognitive principles into AI for enhanced performance and security.