This work proposes an efficient n-party computation framework with malicious security in the outsourced case. We consider a non-symmetrical trust setting that \(n-1\) parties, including \(P_0\) , is semi-honest and one party is malicious. Once the party in \(\{P_1,P_2,...,P_{n-1}\}\) misbehaves, we will caught and exclude it, achieving robustness by executing the further computations among semi-honest parties. Unlike previous works that focused on traditional robustness, we do not need to delegate the computation to an identified honest party, which contradicts the usual expectation of privacy and hinder the adoption in practice. Our framework bridges the gap between semi-honest and malicious honest-majority multi-party protocols, and is well suited for privacy-preserving machine learning. In general, machine learning depends on large amounts of data and powerful computing power, leading resource-constrained clients to resort to secure outsourced computing (SOC) approach. In this work, we utilize the n-party framework to design efficient building blocks for machine learning. We significantly improve the efficiency of blocks and implement machine learning inference in LAN and WAN setting, respectively. Theoretical analysis and experimental results show that our framework outperforms the existing honest-majority protocols, such as SWIFT (USENIX Security’21) in terms of “robustness” and Fantastic Four (USENIX Security’21) in terms of efficiency.