| Popis: |
Data is essential to the performance of artificial intelligence (AI) based malware detection models. System APIs, which allocate operating system resources, are important for identifying malicious behaviors. However, few studies have been conducted on data in the malware detection AI model. They overlooked collection of benign data, which is as important as malware data, and data characterization of system APIs. As an optimization method for data-driven artificial intelligence, this paper studied the data collection, purification, preprocessing, and vectorization for EXE files and system APIs. The objectivity of the data was ensured by using global data, and a more robust model could be created by collecting benign data from Virus Total. By analyzing the weight distribution according to the order of system API execution, we identified that major malicious behaviors occurred at the beginning of execution. We found the optimal API length and optimal dimension (feature number). Finally, accuracy of the N-gram model ranged from 97.62 to 95.73, and that of the Word2Vec model ranged from 97.44 to 95.89. In the generalization performance test using different data from the source of the training ones, we confirmed that N-gram was affected by the quantity of training data, and Word2Vec was affected by data similarity. This study systematized the entire procedure of AI data processing for malware detection, and is the first study to compare and analyze statistical vectors and word embeddings based on the characteristics of system APIs. |
