Semiconductor testing produces high-dimensional sensor data that can benefit from machine learning, but deploying such models in practice raises security concerns. This thesis explores the feasibility of secure deep learning inference in multi-party computation settings, with a focus on predictive performance and runtime efficiency. We implement and evaluate six neural network models using two state-of-the-art MPC frameworks, CrypTen and MP-SPDZ, in total covering six protocols with varying security guarantees. To optimize performance, we integrate dimensionality reduction through principal component analysis, adapt batch normalization, and analyze batching strategies under realistic network constraints. Our results show that while higher security guarantees significantly increase computation time, model size, batching and careful fixed-point representation enable practical runtimes and predictive performance. We find that small and lightweight models achieve the best trade-off between accuracy and efficiency, and that the Brain protocol offers a promising balance between runtime and security. This work provides crucial insights for applying secure machine learning inference in potential industrial semiconductor testing pipelines.