Exploiting epistemic uncertainty of the deep learning models to generate adversarial samples

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial samples, which are carefully crafted input perturbations.
• Current adversarial machine learning research often focuses on model loss functions for attacks and defenses.
• A gap exists in exploring uncertainty quantification for adversarial attack strategies.

Purpose of the Study:

• To explore the use of quantified epistemic uncertainty from Monte-Carlo Dropout for adversarial attacks.
• To develop novel attack strategies by perturbing inputs into shifted-domain regions.
• To exploit model uncertainty in discriminating between original and shifted data distributions.

Main Methods:

• Utilized Monte-Carlo Dropout sampling to quantify epistemic uncertainty.
• Developed a hybrid attack approach leveraging epistemic uncertainty.
• Tested the attack on MNIST Digit, MNIST Fashion, and CIFAR-10 datasets.

Main Results:

• The proposed hybrid attack increased success rates on MNIST Digit from 82.59% to 85.14%.
• Attack success rates on MNIST Fashion improved from 82.96% to 90.13%.
• CIFAR-10 saw an increase in attack success rates from 89.44% to 91.06%.

Conclusions:

• Epistemic uncertainty quantification is a viable method for crafting effective adversarial attacks.
• The proposed method successfully targets DNN vulnerabilities in shifted-domain regions.
• This research contributes a new perspective to adversarial machine learning beyond loss function-based approaches.