With the proliferation of Online Social Networks (OSNs), the emergence of fake accounts has become a significant concern, undermining user trust and spreading misinformation. This comprehensive review delves into the application of machine learning techniques for detecting fake accounts in OSNs, focusing on four principal categories: graph-based, profile-based, behavioral-based, and content-based features. Graph-based approaches leverage the network structure of social connections, analyzing patterns and anomalies in user interactions and relationships. Profile-based methods scrutinize account characteristics, including registration details and personal information, to distinguish genuine users from imposters. Behavioral-based strategies examine user actions, such as frequency and timing of activities, to identify irregularities typical of fake accounts. Lastly, content-based techniques assess the textual and multimedia content shared by users, seeking markers of authenticity or fabrication. This review synthesizes recent advancements and methodologies in each category, evaluating their effectiveness and limitations. We also discuss the challenges of integrating these diverse approaches and the potential of hybrid models that combine multiple feature categories for enhanced accuracy. The paper aims to provide a roadmap for future research in this critical area, highlighting emerging trends and proposing directions for developing more robust and scalable detection systems. This exploration is pivotal for maintaining the integrity and trustworthiness of OSNs and safeguarding users against the pernicious effects of fake accounts.