Design of Software Architecture for Neural Network Cooperation: Case of Forgery Detection

Recent technological advances in media tampering has been the cause of many harmful forged images. Tampering detection methods became major research topics to cope with it in the neural network community. The methods almost always aim at detecting a specific forgery. That is, a general detecting method to find any tampering has not been invented so far. This paper concerns about a software architecture for organizing multiple neural networks to detect multiple kinds of forgeries. The key issue here is to construct, from the meta-level, a mechanism for an ensemble of front-end neural networks to select a neural network which makes a decision. Under this architecture, we implemented a prototype for detecting forged images resulted from multiple tampering methods of copy-move and compression. In order to demonstrate that our architecture works well, we examined a case study with a total of 120,000 patches which consist of three classes of copy-move, compression and untampered data, 40,000 patches for each. The result shows our proposed method successfully classified 108,954 out of 120,000 patches with 90.82 % accuracy. We also give discussions on our architectural implication to avoid concept drift. Our architecture is designed to be a context-oriented and meta-level, which has a two-layered structure: meta and base. The neural networks can be categorized into base-level components, whereas a component coordinating the networks is addressed in meta-level. The architecture explains that the concept drift can be handled in the meta-level. Through the discussions on the techniques of transfer learning, online learning, and ensemble learning in terms of the architecture we constructed, it is concluded that we could construct a universal architecture to coordinate machine learning components.