Associating illumination inconsistencies and deep learning methods to detect image splicing

• Grant number: 18/00858-9
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Effective date (Start): July 01, 2018
• Effective date (End): June 30, 2019
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Tiago Jose de Carvalho
• Grantee: Thales Augusto Paletti Pomari
• Host Institution: Instituto Federal de Educação, Ciência e Tecnologia de São Paulo (IFSP). Campus Campinas. Campinas , SP, Brazil

Abstract

Motivated by the large dissemination of images through the internet, tools that guarantee the veracity of suspicious images are more and more necessary. A fake image can cause problems of unimaginable size and this type of problem can be avoided by developing and deploying new digital forensic techniques for document analysis. For organs like Federal Police and the Judiciary, this kind of methods are essential. The result of an image analysis can easily change the direction of an investigation process, for example. The analysis of an expert based on an appropriate scientific method is essential in the context of a proper judgment by a judge. This project proposes the investigation of a method to perform the detection of image splicing. Our proposal is to develop a method that detects inconsistencies in the illumination of images using illuminant maps, which highlight these types of inconsistencies in fake images and robust architectures of deep convolutional networks in order to detect when and where falsifications of the composition type occur.