Automated Design of Neural Network Architectures With Reinforcement Learning for Detection of Global Manipulations
Chen Yi-fang, Zheng Wang, Z. Jane Wang, Xiangui Kang
Abstract
Deep Convolutional Neural Networks (DCNNs) have been widely used in detection of global manipulations. However, designing effective DCNNs for specific image forensics tasks generally requires domain knowledge and experience gained from abundant experiments, which is time-consuming and labor-expensive. Approaches of automated network designing have been proposed for image classification tasks which are image-content focused, however they may not be suitable to image forensics tasks which rely on identifying subtle traces left by certain image operations. In this paper, we make the first attempt to automate the neural network architecture design for detection of global manipulations. The process of constructing a network is modeled as sequentially selecting optimal architecture modules to generate high-performing CNNs for specific forensic tasks through reinforcement learning. The module-based search space is proposed to make the designing process efficient. Advanced connection patterns (e.g., dense connectivity), which were shown preferred for global manipulation detections, are included in the modules to improve the representational power of the network. Experimental results show that the proposed approach can adaptively construct effective CNN architectures for two common forensic tasks, including multi-purpose forensics and the processing history detection. The auto-designed networks can outperform the state-of-the-art manually designed networks.