Robustness of Sparsely Distributed Representations to Adversarial Attacks in Deep Neural Networks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep learning models excel at tasks but are prone to overfitting and adversarial attacks.
• Dropout regularization is a known technique to improve model generalization and robustness.

Purpose of the Study:

• To investigate the effect of dropout regularization on neural network robustness against adversarial attacks.
• To analyze the relationship between dropout regularization, functional smearing, and adversarial resilience.

Main Methods:

• Examined the impact of varying dropout probabilities on neural network performance.
• Quantified functional smearing, defined as neurons involved in multiple functions.
• Assessed network resistance to adversarial attacks under different dropout conditions.

Main Results:

• Dropout regularization improves resistance to adversarial attacks within a specific probability range.
• Dropout significantly increases functional smearing across various dropout rates.
• Networks with lower functional smearing demonstrate greater adversarial resilience.

Conclusions:

• Dropout regularization enhances robustness to adversarial attacks but is sensitive to dropout probability.
• Reducing functional smearing is crucial for improving adversarial resilience, even with dropout.
• Balancing dropout for generalization and minimizing functional smearing is key for robust deep learning models.