The demand for secure data sharing is growing fast in sensitive domains like healthcare, finance, and IoT. Federated Learning (FL) introduces a decentralised machine learning paradigm whereby models can be trained over distributed nodes without sharing data. Despite its promise, FL faces significant security challenges, such as gradient inversion, model poisoning, and privacy leakage, which involve strong cryptographic techniques. Some cryptographic techniques have been proposed to address potential security concerns in the FL environment. This study provides an overview of the major techniques that include Homomorphic Encryption (HE), Secure Multi-Party Computation (SMPC), Differential Privacy (DP), blockchain integration, and emerging hybrid approaches, and their applications from different perspectives. The techniques, along with their respective strengths and weaknesses, are systematically compared to ensure the identification of appropriate application domains. This review addresses hybrid approaches that combine multiple techniques to achieve an optimal trade-off between privacy, computational efficiency, and scalability. Key challenges such as computational overhead, scalability limitations, and the privacy-utility trade-off are identified, along with notable research gaps in the field. Future directions emphasise on the development of optimised hybrid techniques and strategies to alleviate computational and communication overheads in resource-constrained environments. This study therefore reviews those aspects that might provide useful insights to researchers and practitioners in the development of secure, scalable, and computationally efficient FL systems, and hence facilitate their practical implementation in privacy-sensitive domains.