Feature selection is a critical data preprocessing step in machine learning applications. However, existing privacy-preserving techniques primarily focus on the model training and inference phases, leaving the preprocessing stage as a potential point of privacy leakage. Furthermore, nearly all existing privacy-preserving feature selection schemes are based on the semi-honest security model and cannot withstand attacks from malicious adversaries. To address this critical gap, this paper proposes the first maliciously secure 3-Party feature selection scheme via mutual information (MSFS) to the best of our knowledge. To ensure security and efficiency, we design a series of secure computation sub-protocols, including secure entropy computation, secure MI score computation, secure Top-k score computation, and secure reduction matrix computation. For the proposed protocols, we provide theoretical analysis and security proofs. We also conduct experimental evaluations on both real-world and synthetic datasets. The results show that our scheme can achieve a performance improvement of up to 12.86% while guaranteeing security. This work provides a theoretical and practical foundation for performing secure feature selection in untrusted, collaborative scenarios.