Dottorato in DATA SCIENCE

Titolo della tesi: Media forensics investigations: from the origin to the authenticity of digital content

In recent years, we have observed a massive change in how information is exchanged. On the one hand, the explosion of social media has given birth to a new way of communicating and exchanging news, media, and ideas. On the other hand, the advancement of content manipulation and generation technologies have led to tools capable of recreating incredibly realistic artificial content. All this poses new challenges in verifying the authenticity and integrity of online content. Whenever we come across new media, we must understand its origin, whether it is real or deliberately modified, and verify its authenticity. In this thesis, we will analyze each problem, offering an overview of possible solutions. The first challenge to solve when encountering multimedia content is reconstructing its source. This problem is as essential for verifying online news as for forensic investigations, where an image or video can represent evidence of a crime. Given a media, we wonder if it was captured with a specific offending camera model or if it was instead downloaded from a social platform. Solving this problem means analyzing the compression traces left in the file when it is captured or uploaded to a platform. To solve this challenge, we propose to train neural networks that learn to distinguish these traces, which we define as fingerprints. Specifically, we will show how these fingerprints change from camera to camera and when content is uploaded to a social network, making it possible to reconstruct the source of origin without relying on information such as metadata that can often be modified or deleted. Another significant problem is that of verifying the authenticity of information. Recent advances in the development of artificial intelligence enable the generation of incredibly realistic content: deepfakes. On the one hand, this opens the doors to new applications in entertainment and creativity. On the other hand, it introduces a new generation of super-realistic fake content. The recognition of these contents is possible thanks to a set of factors. First, many of these techniques introduce semantic inconsistencies that are difficult to correct; furthermore, each generative technique leaves specific fingerprints similar to those left by camera models or social media. We will analyze possible strategies for recognizing fake content by exploiting these inconsistencies. All the challenges mentioned so far have one problem in common. Data and information continually evolve, making standard detectors less and less robust as time passes. This is especially true with news, which constantly evolves as events worldwide grow. To prevent this from happening, fake news detectors must continuously learn to classify new information. The last part of this thesis will be dedicated to this topic. On the one hand, we will introduce a continuous learning strategy that allows a detector to learn to classify new news as it is published. Subsequently, we will analyze the vulnerabilities of these techniques concerning a new type of adversary attack. Finally, we will discuss two forensic applications in the fields of ground to aerial matching and insurance.