Robust deep learning object recognition models rely on low frequency information in natural images.

Area of Science:

• Computer Vision
• Neuroscience
• Machine Learning

Background:

• Machine learning models struggle with generalization, particularly vision models vulnerable to adversarial attacks and common corruptions.
• Human visual systems exhibit robustness to these challenges, suggesting potential for bio-inspired solutions.
• Prior research indicates brain-like representations improve model robustness, but the underlying mechanisms remain unclear.

Purpose of the Study:

• To investigate the hypothesis that low spatial frequency preference, inherited from neural representations, contributes to improved machine learning model robustness.
• To explore the role of frequency sensitivity in enhancing the resilience of vision models.

Main Methods:

• Frequency-oriented analyses were conducted, including the creation and use of hybrid images to directly assess model frequency sensitivity.
• Analysis of publicly available robust machine learning models trained with adversarial images or data augmentation.
• Examination of the impact of blurring as a preprocessing defense mechanism.

Main Results:

• Robust machine learning models, including those trained on adversarial data or with augmentation, consistently demonstrated a preference for low spatial frequency information.
• Hybrid image analyses confirmed the link between low-frequency preference and model robustness.
• Preprocessing by blurring improved model defense against adversarial attacks and common corruptions.

Conclusions:

• The low spatial frequency preference inherited from neural representations is a key factor in enhancing machine learning model robustness.
• Leveraging low spatial frequency information offers a promising defense strategy against adversarial attacks and common corruptions in computer vision.
• Blurring, by emphasizing low spatial frequencies, can serve as an effective defense mechanism for robust object recognition.