現今的行動裝置普及，再加上Android作業系統的市占率越來越高，Android惡意程式增長速度越來越快，要如何準確且快速的檢測惡意程式是一個重要的議題。本論文以靜態分析作研究，並且將現今流行的圖像技術應用至Android惡意程式檢測領域中，與現有研究不同的是本研究目標設計出一種有效的分類模型，來解決惡意程式分析上模型的訓練時間冗長的問題。現有圖像惡意程式研究，大多採用VGG Net作為分類器且訓練時間冗長，本研究將自動編碼器(Autoencoder)與圖像領域上使用的深度卷積神經網路(Convolutional Neural Network)結合，運用在惡意程式分析上，旨在縮短訓練時間且達到良好的準確度。自動編碼器（Autoencoder）透過卷積層可以將輸入圖片進行特徵萃取，獲取更低維的向量，此過程可以當作是一種圖像壓縮技術，並提取重要資訊，捨棄不需要的圖像特徵；現今圖像領域中深層卷積模型Efficient Net以較多的卷積層數來獲取圖片更細節特徵，再加上有殘差網路(Residual Network)架構，減少網路退化的問題。本研究採用卷積自動編碼器，並證實可以提取惡意程式特徵將資料集維度縮小，減少訓練時間，並利用Efficient Net作為分類器，在準確度不變的前提下，縮短75%到80%至約500秒的訓練時間。;With the popularity of mobile devices today and the increasing market share of Android operating systems, Android malware is growing faster and faster. How to detect malware accurately and quickly is an important issue. This paper uses static analysis for research, and applies today′s popular image technology to the Android malware detection field. Unlike the existing research, this research goal is to design an effective classification model to solve the problem of lengthy training time and can also improve accuracy. Most of the existing image malware researches use VGG Net as the classifier and they cost lots of time to train. This study combines the Autoencoder and the deep convolutional neural network used in the image field. The malware analysis aims to shorten the training time and achieve good accuracy. Autoencoder can extract feature of input picture through convolutional layer to obtain lower dimensional vector. This process can be regarded as an image compression technology. By extracting important information and discarding unnecessary image features to reduce the dimension. Nowadays in the image field, the deep convolution model Efficient Net uses more convolution layers to obtain more detailed features of the picture, plus a Residual Network architecture to reduce the problem of network degradation. This study uses a convolutional autoencoder and proves that it can extract malware features to reduce the dimension of the data set and reduce training time. under the premise of using different data sets and unchanged accuracy, shorten Up to about 500 seconds of training time.
