<p>Machine learning(ML) has emerged as a fundamental element of innovation in several industries, providing unparalleled powers in data processing, decision-making, and automation. The growing use of ML systems has presented considerable security and privacy challenges, especially in resource-limited contexts such as IoT devices and edge computing platforms. This work examines lightweight security and privacy-preserving solutions designed to mitigate these vulnerabilities, emphasizing both cryptographic and non-cryptographic methods. The work presents a detailed classification of security vulnerabilities aimed at ML systems, encompassing data poisoning, adversarial attacks, model inversion, and training data breaches. It assesses cryptographic methods, including homomorphic encryption and safe multi-party computation, alongside non-cryptographic strategies such as differential privacy, defensive distillation, and federated learning. The work delineates significant obstacles, including processing overhead, adversarial robustness, scalability, and interaction with legacy systems, and proposes specific countermeasures to address these concerns. Additionally, the work integrates empirical evaluations and comparative benchmarks to guide practical deployment and indicates future research areas, highlighting the necessity for scalable cryptographic methodologies, sophisticated adversarial defenses, and interdisciplinary approaches to augment machine learning security and privacy. This article seeks to reconcile stringent security requirements with practical deployment limitations by offering actionable insights and novel methodologies, hence enabling the secure and dependable utilization of ML systems across many applications.</p>