Marija Ivanovska, PhD
Assistant
Laboratory of Machine Intelligence (LMI)
Faculty of Electrical Engineering, University of Ljubljana
Tržaška cesta 25
SI-1000 Ljubljana, Slovenia
E-mail: marija.ivanovska@fe.uni-lj.si
Research opportunities are open to local, exchange, and visiting students. Projects are adapted to the student’s background, interests, and available time. A project may focus on reproducing and analysing an existing method, designing a new experiment, evaluating models across datasets, or developing new algorithms or evaluation approaches.
Students interested in gaining research experience are welcome to contact me to discuss possible projects. Strong project outcomes may be further developed into a thesis, conference paper, or journal publication. Current research opportunities are organized around three main research directions.
Biometric Attacks & Synthetic Media
Modern biometric systems must remain reliable even when images or videos are intentionally manipulated. Projects in this area investigate how attacks are created, how they affect biometric systems, and how they can be detected. Possible research problems include:
- Face morphing attack detection: detecting morphed identity photographs produced using landmark-, GAN-, or diffusion-based methods.
- Deepfake and face manipulation detection: studying visual, frequency, or semantic cues that distinguish authentic from manipulated faces.
- Presentation attack detection: detecting printed faces, replay attacks, masks, or other attempts to spoof biometric sensors.
- Fake ID document detection: identifying manipulated portraits, face replacement, retouching, or other alterations in identity documents.
- Biometric system vulnerability: evaluating how recognition systems behave under previously unseen or emerging attacks.
Synthetic Data for Biometrics
Synthetic data can reduce dependence on sensitive biometric datasets and enable controlled experimentation, but it also raises questions about realism, identity preservation, bias, and security. Possible research problems include:
- Synthetic face generation: generating realistic biometric samples using GANs, diffusion models, or foundation models.
- Synthetic training data for face recognition: studying when models trained on synthetic identities can match systems trained on real data.
- Quality and identity preservation: measuring whether generated samples preserve useful biometric characteristics.
- Privacy-aware biometric learning: investigating whether synthetic data can reduce privacy risks while retaining detection and/or recognition performance.
- Evaluation of synthetic datasets: analysing diversity, realism, demographic coverage, and representation-space properties.
Generalizable & Explainable AI
Many detectors work well on attacks seen during training but fail when the manipulation technique, dataset, or acquisition conditions change. A central research challenge is therefore to develop models that generalize beyond known examples and produce decisions that can be meaningfully interpreted. Possible research problems include:
- Generalizable attack detection: developing detectors that remain effective on previously unseen manipulation techniques.
- Anomaly and one-class learning: learning predominantly from bona fide data when representative attack samples are unavailable.
- Self-supervised learning: exploiting image structure and synthetic artifacts without relying on large manually labelled attack datasets.
- Foundation and vision-language models: investigating whether pretrained multimodal models can detect or reason about biometric attacks with little or no task-specific training.
- Explainable multimodal AI: developing and evaluating methods that make model decisions more transparent and interpretable.