ML security toolbox

Abstract

Critical applications are now incorporating more and more machine-learning (ML) models, this implies that the aforementioned security and privacy flaws must undergo tough screening to prevent attacks. The main objective of this project lies in the creation of a comprehensive se- curity toolbox that will be used for assessing the adversarial robustness of previously trained ML models before their deployment as APIs. The toolbox enables users to configure attack parameters, input datasets, and perform adversarial attacks through the interface. It supports five key attacks: FGSM, I-FGSM, MI-FGSM, Carlini & Wagner (C&W), and PGD, providing comprehensive metrics such as accuracy loss and visualizations of adversarial examples.The per- formance evaluation criteria, such as the degree of accuracy degradation and resilience against adversarial perturbations, offer solutions that are relevant and realistic in a hostile environment of model behavior. By combining cutting-edge adversarial techniques with an intuitive Flask- based platform, the proposed toolbox aims to facilitate the pre-deployment evaluation of ML models, ensuring security and reliability in real-world applications. Furthermore, our imple- mentation emphasizes scalability and adaptability to diverse datasets and model architectures, addressing gaps in existing security evaluation frameworks..