BIDO: A Unified Approach to Address Obfuscation and Concept Drift Challenges in Image-based Malware Detection

To identify malicious Android applications, various malware detection techniques have been proposed. Among them, image-based approaches are considered potential alternatives due to their efficiency and scalability. Recent studies have reported that these approaches suffer significant performance declines when confronted with obfuscation or concept drift. However, existing solutions often treat these two challenges as different problems, offering independent solutions. These techniques overlook the fact that both challenges share a common statistical root, out-of-distribution, and research from this perspective remains limited. In response, we propose BIDO, a hybrid image-based malware detector designed to enhance robustness against both obfuscation and concept drift simultaneously. Specifically, to improve the discriminative power of image features, we introduce a local feature selection module that identifies informative subregions within malware images. Second, to enhance feature robustness, we model pairwise cross-modal dependencies in an outer product space, enabling the extraction of stable co-occurrence patterns. Third, to ensure feature compactness, we design a learnable metric that pulls samples with identical labels closer while pushing apart those with different labels, regardless of obfuscation or concept drift. Extensive experiments on the real-world datasets demonstrate that BIDO significantly outperforms existing baselines, achieving higher robustness against both concept drift and obfuscation. The source code is available at: https://github.com/whatishope/BIDO/.