RobustMap: Visual Exploration of DNN Adversarial Robustness in Generative Latent Space

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Traditional methods for assessing deep neural network (DNN) adversarial robustness provide limited insights, offering only a single score based on a fixed set of test samples.
• This lack of comprehensive evaluation hinders a deep understanding of a DNN's vulnerabilities and defense mechanisms.

Purpose of the Study:

• To develop a novel approach for visualizing the adversarial robustness of deep neural networks (DNNs).
• To enable a more comprehensive understanding of DNN robustness by visualizing its distribution.
• To provide users with tools to explore and interpret DNN robustness effectively.

Main Methods:

• Training a generative model (GM) on existing test samples to create a distribution of DNN robustness over infinite generated samples within the GM's latent space.
• Developing methods to map the GM's high-dimensional latent space to a lower-dimensional plane for effective visualization.
• Designing a predictive network to estimate DNN robustness on large datasets, accelerating the distribution rendering process.
• Creating a user-friendly system for multi-perspective exploration of the DNN robustness distribution.

Main Results:

• The proposed approach successfully visualizes DNN adversarial robustness as a distribution, offering richer insights than traditional methods.
• The generative model approach allows for the extension of test samples, leading to improved feature coverage.
• Experimental results, both subjective and objective, validate the usability and effectiveness of the developed visualization technique.

Conclusions:

• The novel approach provides a comprehensive and intuitive method for understanding DNN adversarial robustness.
• The visualization of robustness distribution enhances the ability to identify vulnerabilities and assess defense strategies.
• The developed system empowers users to explore and analyze DNN robustness from various perspectives, facilitating better model development and security.