Attention 기반 Multimodal 분석 기법을 통한 ELF 악성코드 탐지 성능 향상 방안 연구 = A Study on Improving ELF Malware Detection Performance through Attention-Based Multimodal Analysis

다국어 초록 (Multilingual Abstract)

Recently, multimodal deep learning techniques that leverage different types of data to learn complementary information have gained significant attention in various artificial intelligence applications. In this study, we propose a multimodal malware detection model that extracts features from executable files by inputting grayscale images into a CNN and entropy-based block sequence data into an RNN. These features are then dynamically integrated using an attention mechanism that assigns importance weights to each feature. Existing Early Fusion and Late Fusion methods suffer from limited interaction between features. In contrast, the proposed model addresses this limitation by applying an attention-based weighting strategy, thereby effectively enhancing detection performance. In addition, to address the data imbalance issue and objectively evaluate the model’s performance, precision, recall, and F1-score were used as the primary evaluation metrics. Experimental results show that the model achieved superior performance compared to the existing methods, recording a precision of 97.81%, a recall of 95.04%, and an F1-score of 96.40%. These results empirically demonstrate that the proposed approach can effectively classify and detect malware with diverse characteristics.