News from the Department of Audiovisual Analysis and Biometric Systems
The team participated in prestigious summer schools, presented research findings at international conferences, and developed their own projects, which are already being applied in the fight against disinformation and synthetic content.
One of the most significant highlights was a series of specialized trainings. Adam Baran attended the International Summer School for Advanced Studies on Biometrics for Secure Authentication in Italy, where the main theme was deepfakes and presentation attacks – topics that form the foundation of his doctoral research. Michał Koźbiał took part in the Oxford Machine Learning Summer School, exploring multimodal models and adaptive AI agents. Meanwhile, Asia Gajewska and Michał Ołowski spent the summer in Tuscany at the IEEE–EURASIP Summer School, where they studied methods of continual and adaptive learning as well as the use of foundation models in niche problems. Two new PhD students joined the department – Adrian Kordas and Donat Stankiewicz – while Michał Banach successfully defended his master’s thesis.
The team also actively participated in international conferences. In Rotterdam, Asia Gajewska and Alicja Martinek presented a paper on the impact of speech fluency style on the vulnerability of voice verification systems. In Toulouse, Alicja Martinek and Ewelina Bartuzi-Trokielewicz, in collaboration with the Bydgoszcz University of Science and Technology, presented research on detecting disinformation through text analysis. In the autumn, Alicja Martinek presented findings on speech generation artifacts in South Korea, where she and Ewelina Bartuzi-Trokielewicz met with representatives of the AI Safety Institute. Shared challenges in deepfake detection pave the way for fruitful scientific collaboration.
Research conducted at the Department covers both audio analysis and computer vision. Michał Ołowski explores ways of using noise to improve model robustness, Asia Gajewska investigates catastrophic forgetting in continual learning, and Alicja Martinek develops methods for audio source tracking using prototypical networks. Donat Stankiewicz works on detecting partial-deepfakes, recordings that combine authentic and synthetic elements. In the field of vision, Adrian Kordas develops the VLLM classifier, Adam Baran studies PADs in a multichannel approach, and Michał Koźbiał refines algorithms for detecting illegal content. Ela Gomulska analyzes the misuse of logos in deepfakes, while Jarosław Wójtowicz and Michał Banach expand training datasets, including those related to lip-sync technologies.
In summary, 2025 was a year of dynamic growth and scientific achievements for the Department. Participation in summer schools and conferences enabled the team to broaden its expertise and establish international collaborations. The DROZD project and research on multimodal artificial intelligence demonstrate that combating disinformation and deepfakes requires an interdisciplinary approach, combining audio, image, and text analysis. The team eagerly awaits the publication of their results in renowned journals, while winter evenings are devoted to further experiments and reading the latest works in the field of artificial intelligence.