Analysis and detection of synthetic audiovisual materials - deepfake

Deepfake technology, thanks to the rapid advances in artificial intelligence and related generative techniques, is becoming a tool that anyone can access. It is used to modify and create new audio-visual content depicting a selected person, thereby putting them in a situation that may never have happened. Cybercriminals use the victim’s voice and image to phish for money, create discrediting content or to blackmail. Often public figures become the victims, and the generated content, using their image, is designed to persuade people to make risky investments, register on fake cryptocurrency exchanges, or purchase virtual goods. Deepfake using public figures can also be disinformative. Such actions can destabilize security, or discredit the person internationally.

Our project involves detailed analysis of audiovisual deepfakes. It is focused on developing advanced tools and methodologies for recognizing facial manipulation, analyzing synthetic voices, and detecting artifacts, using machine learning, deep networks, and biometrics techniques. The goal of the project is to create a comprehensive, multi-modal system capable of effectively identifying materials modified using deepfake technology, which is key to ensuring digital security and protecting against disinformation. The project is divided into several major phases:

Facial image manipulation recognition – we are developing AI and machine learning algorithms that can identify inauthentic alterations in images and videos, with a particular focus on facial manipulation. These modules take into account both static and dynamic facial features.

Synthetic voice analysis – we apply signal processing and machine learning techniques to distinguish natural human voices from synthetic ones, with the development of methods for evaluating voice biometric features and their changes due to manipulation, including in the context of Polish.

Artifact detection – we are developing methods to identify irregularities and anomalies in audiovisual materials that may suggest the use of deepfake technology, such as inconsistencies in skin texture, unnatural blinking, or artifacts around the mouth.

Facial biometrics and speaker verification – we implement advanced biometric techniques to analyze and verify people’s identities based on facial and voice features, with integration of identity verification systems with other project modules for improved accuracy of deepfake detection.

Model performance analysis – we use model performance explainability techniques and data-centric approaches to help understand how algorithms make their decisions, which is key to improving their effectiveness and user confidence. These can be used in evidence analysis.

Education and dissemination – we participate in workshops, meetings and trainings for various groups to educate about deepfake, raise awareness about the risks of deepfake technology.

Do you want to implement a project with us? Meet for a joint workshop?

Contact us: deepfake@nask.pl