- 學位論文
基於CNN的假圖片分類模型改善與評估
Improvement and Evaluation of CNN-based Fake Image Classification Model
摘要
在當今的社會中，人們可以輕易地透過網路獲得資訊，而最常見的一種方法即是透過圖片的傳播。然而，隨著圖片編輯軟體的快速發展，人們可能沒有察覺到自己已經被偽造的圖片所欺騙。因此，我們將提出一個機器學習的方法，以實作出虛假圖片的偵測。在此研究中，會使用CASIA此資料集做訓練，其中包含了兩種不同的偽造圖片：圖片拼接（Splicing）以及圖片複製移動（Copy-move）。 基於卷積神經網路，我們會使用局部二值模式（Local Binary Patterns）和錯誤級別分析（Error Level Analysis）這兩種資料前處理方法去提升模型的準確度。根據實驗結果，自建的CNN搭配ELA方法表現最好，準確度達到約90%。此外，模型對於假圖片的敏感度也超過了95%。為了了解模型的弱點，我們同時也對預測錯誤的圖片進行了分析，並在最後討論此研究方法的限制為何。
並列摘要
Nowadays, people can easily acquire information on the Internet, and the most common way to deliver messages is through the images. However, with the rapid development of editing software, people might be cheated by the tampered images without consciousness. Therefore, we will propose a machine learning method to implement fake image detection. In this study, the used dataset is CASIA 2.0 which include two types of fake images: splicing and copy-moving. Based on Convolutional Neural Network (CNN), and we will additionally apply the data pre-process methods, Local Binary Patterns (LBP) and Error Level Analysis (ELA), which are empirically effective in image forensic to increase the accuracy of CNN models. The best model of the research is a self-build CNN with ELA, and the accuracy is about 91% in training and 90% in validation. Furthermore, the sensitivity towards fake images is over 95%, which is well-performed model on fake image detection. Meanwhile, we analyzed the false predictions of the model to realize the weakness of the proposed model. Last, we also discuss the limitations of this research method.