Detecting Adversarial ML Threats in Agile Workflows : A Cybersecurity Lens

The use of machine learning (ML) in agile software development is expanding, creating new cybersecurity risks, especially adversarial ones, which target weaknesses in its models. Adversarial ML attacks are the formulation of inputs that can deceive or hack ML models and frequently have small perturbations that are unnoticeable by humans yet devastating to algorithmic decision-making due to small perturbation. Such vulnerabilities may pass unnoticed in environments where speed, iteration and continuous delivery are paramount, so limited time will be available to perform tests and validate at a higher level. This paper explores finding and dispelling adversarial threats in ML, which started with a security-based approach, i.e., agile workflows. We look at the real-time application of adversarial attacks in the financial, content moderation, and autonomous systems. The analysis performed on ready-made attack algorithms (e.g. FGSM, PGD), in the peculiarities of the simulation, proves the vulnerability of models used in ordinary practice and the efficiency of defense strategies, such as adversarial training. The effect of possible attack vectors and modelling defenses are depicted by visualization. Some of the significant issues we also find in the security of ML in the agile environment are adversarial threat modelling, little security knowledge among development teams, and poor testing infrastructure. Lastly, we suggest remedies to the current state of affairs and how to integrate strong measures of adversarial defense and resilience within agile MLOps pipelines whilst maintaining development speed.