Image Scaling Attacks on Machine Learning Algorithms: A Cyber Security Perspective

The integrity of machine learning (ML) models is increasingly threatened by adversarial attacks that exploit vulnerabilities in the preprocessing stages. This research explores the susceptibility of ML algorithms to image scaling attacks, a type of adversarial attack that manipulates the input images during resizing operations to induce incorrect model predictions. Focusing on traffic sign recognition and utilizing the German Traffic Sign Recognition Benchmark (GTSRB) dataset, we investigate the resilience of neural network-based classifiers against these attacks. By employing a Keras-based ML model, we examine the impact of image scaling attacks on the accuracy and reliability of traffic sign recognition systems, which are critical for autonomous vehicle navigation and traffic management. This study details the process of preprocessing image data, applying various scaling techniques, and developing attack strategies that subtly alter traffic sign images. Through experiments with different neural network architectures, including Convolutional Neural Networks (CNNs), we reveal the vulnerability of these models to carefully designed image manipulations and propose mitigation strategies to enhance their robustness. Our research provides valuable insights into ML vulnerabilities and advocates for a framework to develop more secure ML systems, emphasizing the importance of understanding and mitigating such attacks to ensure the security and dependability of ML models.