Adversarial training and attribution methods enable evaluation of robustness and interpretability of deep learning...

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep learning models excel but exhibit fragility to adversarial attacks and out-of-distribution data.
• Model interpretability is crucial for understanding and addressing this fragility.

Purpose of the Study:

• Investigate the impact of adversarial training on input attribution methods.
• Benchmark and identify reliable input attribution techniques for deep learning models.

Main Methods:

• Combined adversarial and input attribution approaches for image classification.
• Evaluated signal-to-noise ratio of input attribution and correlated with model confidence.

Main Results:

• Adversarial training yields distinct input attribution matrices compared to standard methods.
• Identified reliable input attribution approaches and confirmed adversarial training enhances prediction robustness.

Conclusions:

• Adversarial training improves deep learning model robustness and interpretability.
• The proposed methodology enhances confidence in deep learning model reliability and is extensible to other domains.