SMCD: Privacy-preserving deep learning based malicious code detection

With the rapid development of the Internet, malicious code has been continuously exposing security issues, posing a significant threat to people’s online lives. Deep learning has shown significant impact in the field of malicious code detection, multiple providers of malicious code data can offer more diverse data for deep learning, thereby improving the accuracy of malicious code detection models. However, this may raise privacy and security concerns regarding the training data and models. To address this challenge, our paper introduces an advanced, secure deep learning framework collaboratively trained across multiple parties. We first use privacy set intersection techniques to align the provided malicious code data from the participants, ensuring that they have the same attributes. The aligned data from each data provider is then securely shared with three cloud servers through secret sharing. The three cloud servers implemented a secure model training process through secure multiparty computation. Our experiment demonstrates that our secure malicious code detection protocol exhibits satisfactory performance.