Resilience Comparison of CNN-enabled Systems under Adversarial Attacks on Different Computing Environments  
Author Karen Da Mata

 
Co-Author(s) Priscila Silva;  Fatemeh Salboukh; Lance Fiondella

 
Abstract Convolutional Neural Networks (CNN) are widely used in applications such as image classification and autonomous systems but are vulnerable to adversarial attacks, requiring adaptation to dynamic and hostile conditions while maintaining high performance. Previous studies have considered the impact of different Python frameworks and hyperparameters on CNN model performance. However, fewer studies have considered the role of computing environments in resilience under attack scenarios. This study compares a CNN resilience model in adversarial situations in three different computing environments, including MacOS, Linux, and Windows, which are composed of diverse hardware elements. An image classification model is subjected to two generative adversarial attacks and iteratively retrained to improve system performance whenever it falls below a pre-specified threshold. Data related to performance is collected from each computer and then employed to compute several resilience metrics that quantify the computer’s ability to preserve the CNN model’s performance. The results indicate that the Linux machine outperformed the other two platforms, preserving performance by 2.7% more and reducing performance loss by 16.16%. Thus, the Linux system was more suitable for running the CNN-enabled application under adversarial conditions, enabling greater resilience.

 
Keywords Resilience engineering, Machine learning resilience, Convolutional neural network, Adversarial attacks, AI test and evaluation
   
    Article #:  RQD2025-47
 

Proceedings of 30th ISSAT International Conference on Reliability & Quality in Design
August 6-8, 2025