Luca Guarnera
University of Catania, Research Fellow (RTD-a)
email: luca.guarnera@unict.it

Alessandro Ortis
University of Catania, Assistant Professor (RTD-b)
email: ortis@dmi.unict.it











Adversarial Machine Learning on Multimedia Forensics Domain



Abstract

In recent years, Artificial Intelligence (AI) algorithms based mainly on the use of deep neural networks have achieved considerable success in various contexts, especially in Computer Vision applications. These technologies are often applied in dangerous contexts, such as in the pornography industry, having as main objectives denigrating a person, gaining success. The high performance achieved by modern generative architectures such as Generative Adversarial Networks (GAN) and Diffusion Models (DM) have drawn the attention of researchers to emerging issues related to the authenticity and integrity of multimedia content and the vulnerability of AI-based applications. The advent of Adversarial Machine Learning algorithms has led to the emergence of several approaches able to compromise the proper performance of machine learning algorithms, causing misclassification. Adversarial Machine Learning techniques combined with generative models' technology turn out to be extremely powerful and especially dangerous tools if used for illicit purposes. The creation of new forensic algorithms capable of defining the authenticity and integrity of multimedia content turns out to be necessary. The tutorial will provide an introduction to Adversarial Machine Learning and Generative Models with applications in various contexts.

Tutorial description

In recent years, Artificial Intelligence (AI) algorithms based mainly on the use of deep neural networks have achieved considerable success in various contexts, especially in Computer Vision applications. These technologies are often applied in dangerous contexts, such as in the pornography industry, having as main objectives denigrating a person, gaining success. The high performance achieved by modern generative architectures such as Generative Adversarial Networks (GAN) and Diffusion Models (DM) have drawn the attention of researchers to emerging issues related to the authenticity and integrity of multimedia content and the vulnerability of AI-based applications. The advent of Adversarial Machine Learning algorithms has led to the emergence of several approaches able to compromise the proper performance of machine learning algorithms, causing misclassification. Adversarial Machine Learning techniques combined with generative models' technology turn out to be extremely powerful and especially dangerous tools if used for illicit purposes. The creation of new forensic algorithms capable of defining the authenticity and integrity of multimedia content turns out to be necessary. The tutorial will present how various generative models such as GANs and DMs can be used to generate so-called synthetic media or commonly referred to as "deepfakes," which can be used for malicious applications, as well as methods and best practices for detecting deepfake content. The tutorial will also provide an introduction to Adversarial Machine Learning and the main techniques for specific attacks (e.g., CEO-phishing fraud) on any type of media content. A practical part will focus on the use of modern generative synthetic data creation solutions, development of countermeasures, and Adversarial Machine Learning applications in this context. In the first part (1-2 hours) Dr. Alessandro Ortis will present an overview of Adversarial Machine Learning and about GAN and models used to create deepfakes contents, providing a series of state-of-the-art papers and examples, focusing on the reasons why deepfakes are a serious problem. Then, in the last part (1-2 hours), Dr. Luca Guarnera will show and explain some state-of-the-art deepfake detection methods and (Practical session) how to easily set up GAN and DM models for the creation of deepfakes and how to detect fake contents.

Syllabus: