In the artificial intelligence era, data-driven computation tasks, such as machine learning, have been playing an essential role as the decision-maker to unlock the value of big data in many fields. Moreover, the ultimate goal of pursuing accuracy and efficiency improvement to better promote the application of data-driven computation has never changed. Since the main method to improve the accuracy of these tasks (e.g., the model training of machine learning) is to increase the diversity of datasets, this requires multiple data providers to share their data. However, data providers, e.g., private companies, are reluctant to share their datasets directly, considering the privacy protection of user information and the leakage prevention of their business secrets. Therefore, how to securely and efficiently perform joint datasets based data-driven computation tasks has become the main problem. In this work, without getting the aid of any trusted-third party (e.g., the cloud server), we construct an efficient maliciously secure two-party mixed-protocol framework for data-driven computation tasks. In particular, we construct a new cryptography gadget called committed oblivious linear evaluation (C-OLE) based on the homomorphic commitments in the malicious model. Then we construct two types of share conversion protocols in the malicious model with the above C-OLE gadget to construct the two-party mixed-protocol framework for data-driven computation tasks. Without utilizing the random oracle, our framework can provide a stronger security guarantee than the other two-party mixed-protocol frameworks in the literature. Furthermore, we conscientiously evaluate the theoretical efficiency of the two shares conversion protocols and provide the result as an important reference for future developers who intend to securely and efficiently instantiate the data-driven computation tasks (e.g., privacy-preserving machine learning applications) in the malicious model.

在人工智能时代，机器学习等数据驱动的计算任务作为决策者发挥着至关重要的作用，可以在许多领域释放大数据的价值。而且，追求准确性和效率提升以更好地促进数据驱动计算的应用的最终目标从未改变。由于提高这些任务的准确性（例如机器学习的模型训练）的主要方法是增加数据集的多样性，这需要多个数据提供者共享他们的数据。然而，考虑到用户信息的隐私保护和商业秘密的泄露，数据提供者，例如私营公司，不愿意直接共享他们的数据集。所以，如何安全高效地执行基于联合数据集的数据驱动计算任务已成为主要问题。在这项工作中，我们在没有得到任何可信第三方（例如，云服务器）的帮助的情况下，为数据驱动的计算任务构建了一个高效的恶意安全的两方混合协议框架。特别是，我们基于恶意模型中的同态承诺构建了一个新的密码学小工具，称为承诺不经意线性评估（C-OLE）。然后我们利用上述C-OLE gadget在恶意模型中构建两种类型的共享转换协议，以构建用于数据驱动计算任务的两方混合协议框架。在不使用随机预言机的情况下，我们的框架可以提供比文献中其他两方混合协议框架更强的安全保证。