Full Program

Implementation and Experimental Evaluation of Defense Techniques against Adversarial Attacks

Summary:

Deep Learning is a key approach that relies on multiple layers of processing to extract patterns and representations from data. Neural Networks are a largely used family of deep learning model architectures that are inspired by the structure of the human brain in order to handle complex data analysis tasks with remarkable efficiency. Despite their advantages, they are susceptible to adversarial attacks — small, often imperceptible alterations to input data that can lead to incorrect predictions. This security issue is especially critical to anomaly-based Intrusion Detection Systems, which rely on neural networks to detect zero-day attacks.

In this paper, we analyzed both adversarial attack methods and corresponding defense strategies, implementing several state of the art techniques. The results of this analysis show pros and cons of the considered defense strategies, allow to identify the more effective ones, and highlight the need for continuous improvements in order to build more resilient models.

Author(s):

Giovanni Maria Cristiano    
University of Naples "Parthenope"
Italy

Salvatore D'Antonio    
University of Naples "Parthenope"
Italy

Jonah Giglio    
University of Naples "Parthenope"
Italy

Giovanni Mazzeo    
University of Naples "Parthenope"
Italy