Image Splicing Detection based on Deep Convolutional Neural Network and Transfer Learning

Abstract

With the easy availability of numerous low-cost and ready-to-use image editing software, nowadays, the forgery of the digital image has become very common and widespread. The fraudsters can deliberately use the tampered images for various malicious activities. Among different types of digital image manipulation, probably the most popular and effective technique is the splicing of images, which makes a natural-looking composite image by cutting and combining different portions from multiple source images. The detection procedure of image splicing can be categorized into two types, which are - feature engineering and machine learning-based detection, and deep learning-based detection with automatic feature extraction. Usually, deep learning-based mechanisms are robust, and they provide high accuracy. However, the deep learning model requires a vast amount of data for training, and it is also time-consuming and costly for its structural complexity. Therefore, in this work, an effective image splicing detection scheme based on deep learning and transfer learning has been proposed that also overcomes the issues of high training time and complex modelling structure. In this work, a Deep Convolution Neural Network (CNN)-based model has been developed where the initial convolution layers are replaced by a pre-trained CNN model MobileNetV2. Instead of training the developed CNN model from the beginning, a transfer learning technique has been adopted to skip the time-consuming initialization of the network. With the help of experimental results, it is shown that our simplified model can successfully detect spliced images with state-of-the-art accuracy despite removing the requirements of very high training data and time.