Abstract:

The use of artificial intelligence to create synthetic media, either from scratch or by changing parts of an existing content, has become very popular. The drivers behind this popularity are in a widespread access to often initially open source deepfake and Generative AI algorithms, but also in the low cost, their performance and their ease of use by nonprofessionals. Some see this trend as a profound paradigm shift that will redefine the foundations of creativity and opens doors to new opportunities in entertainment, gaming and art to mention a few among many applications. But at the same time, several challenges have surfaced, among which concerns about their use in misinformation and fraud. 

In this talk, we will focus on synthetic media in form of visual information and start with an overview of the most popular algorithms in deepfake and Generative AI applied to visual content, highlighting their performance in generating synthetic content through a number of concrete illustrations. We then move to an in-depth analysis of challenges that arise not only in terms of performance to generate synthetic media but also other relevant issues such as how to identify synthetic media as opposed to those generated by sensors operating in the real world, as well as various trust and security issues surrounding them. The talk will conclude by highlighting the current trends in solutions to cope with challenges in synthetic media and in particular two major approaches in their trust, namely, detection and provenance.