Authors - Avni Tyagi, Suman Madan Abstract - Machine Learning Operations (MLOps) has become a paradigm necessary to simplify machine learning systems development, implementation, and operation. Although MLOps focuses on automation, scalability, and fast deployment based on CI/CD practices, issues of security are usually under-explored, making ML pipelines very vulnerable.To examine the main security risks of contemporary ML pipelines, the paper explores the intersections between adversarial machine learning and MLOps and DevSecOps. It determines key attack vectors, such as data poisoning, model tampering, and infrastructure-level exploits, which may impair data integrity, model reliability, and system trustworthiness, through a review of recent literature (2020-2026).It also analyzes mitigation measures like adversarial robustness testing, cryptographic model signing, and continuous monitoring models and looks at new frameworks like SecMLOps and MLSecOps that help to put security in the ML lifecycle.It points out trade-offs between improved security, system performance, and complexity, and the importance of balanced architectures. Results show that adversarial testing and verifying the model with secure artifacts can decrease model failure rates by 3060 percent, and that continuous monitoring can improve the latency of anomaly detection by almost 40 percent.The paper ends with description of future research directions such as standardized benchmarks, enhanced robustness testing, and hardware-aided security of robust AI systems.
Open Zoom