基于去噪扩散隐式模型（DDIM）增强算法的恶意软件疫苗库
收藏资源简介：
随着移动智能终端的广泛普及，Android平台成为全球最大的移动操作系统，带来了丰富的应用生态，也引发了日益严重的网络安全威胁，尤其是恶意软件的快速蔓延。高质量、结构化的恶意软件数据集对于数据科学和网络安全领域具有重要意义，能够为特征提取、模型训练、算法评估等环节提供坚实的数据基础，推动智能检测与防护技术的发展[1][2]。 在现有研究中，CICMalDroid2020数据集由加拿大网络安全研究所（CIC）与加拿大互联网注册局（CIRA）联合发布，作为当前Android平台较为全面的恶意软件数据集，广泛应用于学术研究、恶意软件检测、分类建模及性能评估等场景[3][4]。该数据集覆盖广告软件、银行恶意软件、短信恶意软件、风险软件和良性应用等典型类别，具备良好的代表性和科学性。然而，随着检测算法的不断发展，传统数据集在样本规模、类别均衡和多模态表达等方面的局限性逐渐显现，推动了数据增强和多维特征表达技术的研究。 本研究基于CICMalDroid2020原始数据，通过系统的数据清洗、特征缩放、流形降维、概率扩散合成和GAF图像化等多阶段处理，构建了结构丰富、样本多样、适用于深度学习的Android恶意软件图像数据集。该数据集不仅能够支持新一代智能检测算法的开发，也具备高度的重用价值，适用于模型训练、算法评测、特征可视化等多种数据科学应用场景，为学术界和工业界提供了重要的基础资源。数据集共包括6个核心数据文件，其中：(1) new_data290.csv是清洗后的CICMalDroid2020数据，数据量7.16MB；(2) processed_sqrt_8000.csv 是特征缩放后的数据，数据量13.4MB；(3) UMAP_malware_data128.csv是特征降维后的数据，也是模型训练数据，数据量13.4MB；(4) image_DDIM是扩充后的数据经过GAF转换后得到的图像文件夹， 数据量504MB；(5)DDIM.py是核心算法脚本，数据量8.02KB；(6) requirements.txt 是环境依赖清单，数据量 0.1MB。
With the widespread proliferation of smart mobile terminals, the Android platform has become the world's largest mobile operating system, bringing a rich application ecosystem while also giving rise to increasingly severe cyber security threats, especially the rapid spread of malware. High-quality, structured malware datasets are of great significance to the fields of data science and cybersecurity, as they can provide a solid data foundation for processes such as feature extraction, model training, and algorithm evaluation, and promote the development of intelligent detection and protection technologies [1][2]. In existing research, the CICMalDroid2020 dataset was jointly released by the Canadian Institute for Cybersecurity (CIC) and the Canadian Internet Registration Authority (CIRA). As one of the most comprehensive malware datasets for the Android platform to date, it is widely applied in scenarios such as academic research, malware detection, classification modeling, and performance evaluation [3][4]. This dataset covers typical categories including adware, banking malware, SMS malware, riskware, and benign applications, and has good representativeness and scientific rigor. However, with the continuous development of detection algorithms, the limitations of traditional datasets in terms of sample size, class balance, and multimodal representation have gradually become apparent, promoting research on data augmentation and multi-dimensional feature representation technologies. This study constructs a richly structured, diversely sampled Android malware image dataset suitable for deep learning based on the original CICMalDroid2020 data, through multi-stage processing including systematic data cleaning, feature scaling, manifold dimensionality reduction, probabilistic diffusion synthesis, and GAF-based imaging. This dataset not only supports the development of a new generation of intelligent detection algorithms, but also has high reusability, and is applicable to various data science application scenarios such as model training, algorithm evaluation, and feature visualization, providing an important foundational resource for both academic and industrial communities. The dataset includes a total of 6 core data files, which are detailed as follows: 1. `new_data290.csv`: Cleaned CICMalDroid2020 data, with a size of 7.16 MB; 2. `processed_sqrt_8000.csv`: Data after feature scaling, with a size of 13.4 MB; 3. `UMAP_malware_data128.csv`: Data after feature dimensionality reduction, also used as model training data, with a size of 13.4 MB; 4. `image_DDIM`: Image folder obtained by converting augmented data via GAF transformation, with a size of 504 MB; 5. `DDIM.py`: Core algorithm script, with a size of 8.02 KB; 6. `requirements.txt`: Environment dependency list, with a size of 0.1 MB.
