Unimore leads DeepFake research with CoDE: the most advanced system to distinguish the real from the fake
DeepFakes, a term combining ‘Deep Learning’ and ‘Fake’, are images, videos and sounds generated or manipulated by artificial intelligence systems with such precision that they are difficult to distinguish from the real thing. For images, technologies such as StableDiffusion, MidJourney, DALL-E and many others are now within everyone's reach and allow the creation of artificial visual content, often used for industrial, medical, artistic or educational purposes, but also with the risk of manipulating information in a misleading way.
On this aspect, the University of Modena and Reggio Emilia is at the forefront with the creation of CoDE (Contrasting Deepfakes Diffusion via Contrastive Learning), an advanced artificial intelligence system that currently represents the most accurate technology in the world for DeepFake recognition.
CoDE was officially presented at the European Conference on Computer Vision 2024 in Milan, one of the most prestigious international events in the field of computer vision, which was attended by over 5,000 researchers from all over the world. This system was developed by two Unimore PhD students, Lorenzo Baraldi and Federico Cocchi, under the supervision of Rita Cucchiara and Lorenzo Baraldi (namesake) of the ‘Enzo Ferrari’ Department of Engineering in Modena and Marcella Cornia of the Department of Education and Human Sciences in Reggio Emilia.
The CoDE system is based on a contrastive learning architecture, trained not only to discriminate true from false images, but also capable of working on parts of images or pixels processed by imaging tools, to make recognition also robust to voluntary or involuntary transformations of the images themselves, when they are compressed, transmitted, published. The heart of the project is actually massive learning, made possible by extensive work from more than two million real photographs with associated text descriptions. From these starting images, the Leonardo S.p.A. researchers, using their Da Vinci supercomputer, generated more than 9 million artificial images, a total of 3,200 hours of GPU processing, corresponding to about ten years of calculation on a traditional workstation.
These images generated by different systems represent a zoo of possible different generative models and are a fundamental asset for the training of detection systems capable of distinguishing real from artificial images. Thanks to this database, CoDE achieved extraordinary results: an identification accuracy in benchmarks of more than 97%, which is well above the 60% that a human being could achieve. A technology that, although extraordinary, is still not without its limitations: CoDE, in fact, may have difficulty recognising images that have been manipulated at a later stage, with heavy compression or digital reworking, and it is not yet known how well it is able to generalise over all generative models developed in the last two or three years. Moreover, the continuous evolution of image generators makes it necessary to constantly update the system to ensure its effectiveness even in the face of new threats.
The issue of image manipulation does not only concern the academic or scientific world. The DeepFake phenomenon has direct consequences on society, with huge risks of misinformation and falsification affecting both the public and the private sector. Also to address these issues, ELSA (European Lighthouse on Security and Safety AI), a European strategic project that aims to develop new scientific research results for security in the age of artificial intelligence, has been active since 2023. Unimore and Leonardo S.p.A. play a leading role in this project, with the aim of creating AI systems capable of identifying manipulated content.
As part of the ELSA project, Unimore and Leonardo organised an international competition in which numerous research centres from all over the world participated. Thanks to this initiative, millions of images were produced and analysed, which enabled the development of increasingly refined algorithms for the recognition of false images.
CoDE not only identifies an image as true or false, but also provides an assessment of the percentage of reliability of the result and a representation of the space in which the image lies. In fact, the system displays a graphical map indicating whether the image belongs to the real content space or the DeepFake space, also providing indications as to which generative AI system produced the visual content.
The big problem, pointed out by the Unimore research team, lies in the definition of ‘true’. All images, even those captured by cameras or smartphones, reflect a vision of reality filtered by the photographer's gaze or the technologies used to produce it. Even the works of great photographic artists, while representing stylised interpretations of the world, are recognised as real by CoDE. In contrast, images generated completely or partially by AI tools such as Photoshop or StableDiffusion can represent a fictitious reality, albeit visually indistinguishable.
“The definition of fake is now widely used, but it is often reductive and misleading,” says Prof. Rita Cucchiara, director of Unimore's AI Research and Innovation Centre. “It would be preferable to talk about artificial or synthetic images, because they are not just fake content, but creations generated through artificial intelligence algorithms that have extraordinary potential. Think, for instance, of the synthetic renderings used by architects and designers to visualise projects still under development, or the synthetic images used to train diagnostic systems in the medical field, capable of recognising rare diseases through the analysis of visual models. These are tools that, besides facilitating research and development work in many areas, can significantly improve the predictive capabilities of advanced technologies.”
“However,” Prof. Cucchiara concludes, “we must be aware of the manipulative potential that accompanies these technologies, which is why research into the transparency and reliability of AI-generated visual content is just as important as the development of the technologies themselves.”
Faced with these problems, the development of regulatory instruments that, for instance, impose watermarking (a ‘watermark’ inserted in the generated data), while a welcome measure, may not be sufficient to guarantee effective protection against the dissemination of false images. Unimore research team, in collaboration with other European partners, is engaged in the creation of increasingly robust and certified technologies capable of complementing generative artificial intelligence with equally powerful detection and defence tools.
The goal is to make these recognition systems accessible to all, so that anyone can protect themselves against digital manipulation and protect their rights, including copyright.
Categorie: International - english, Notizie_eng
Articolo pubblicato da: Ufficio Stampa Unimore - ufficiostampa@unimore.it