We encounter forged image on a daily basis. From image enhancement on Instagram, Snapchat, etc. to more sophisticated forgeries done in an advanced platform like Photoshop and GIMP, image editing is no longer a professional game. Most forgeries are clear even to the naked eye but with few hours of YouTube tutorials, one can fool most people with his forgery skill.
The problem in hand results from the combination of two things. First, people’s lack of integrity, which is out of this work scope and our focus, and the second is that our world is still founded on the phrase “A picture is worth a thousand words”. From a trail outcome to a newspaper article, one image has a great power in the modern world.
The recent advancement in machine learning, and particularly in deep learning allowed the emergence of new ways to detect forgeries [1][2][3][4], where the main tool was convolutional neural networks. But, none of them provides the necessary accuracy to be admissible in court.
Our proposed solution is a voting schema. Our schema implements an aggregation method on the results of a few state-of-the-art image forgery detection deep learning methods in order to localize the location of the forgeries in images. More specifically, we implemented several methods from recent research and trained several CNN and an aggregation detector.
The results show that our current scheme is still not accurate enough to be used in the real-world. But it is also showing the potential in our proposed solution. Our schema performed better then each of the single methods that composed it.