A complete list of my research publications, primarily in the areas of AI, computer vision, and media integrity.
36 publications
36 publications found
Jonas Dolezal, Sawood Alam, Mark Graham, Maty Bohacek
Measures the prevalence of AI-generated text on the web and finds reduced semantic diversity and increased positive sentiment, but no statistical support for decreased factual accuracy or stylistic homogenization.
Maty Bohacek, Thomas Fel, Maneesh Agrawala, Ekdeep Singh Lubana
Method for identifying blindspots in text-to-image models: concepts the model was trained on but cannot generate.
Sarah Barrington, Maty Bohacek, Hany Farid
Large-scale dataset of real and deepfake footage for detecting face-swap and lip-sync deepfakes.
A. A. McGuire, Maty Bohacek, Hany Farid, Paul Taylor, Sophie J. Nightingale
Studying human perception of AI-generated faces and their realism.
Raina Panda, Daniel Fein, Arpita Singhal, Mark Fiore, Maneesh Agrawala, Maty Bohacek
Interpretable style in vision models using sparse autoencoders for controllable style transfer.
Maty Bohacek, Nino Scherrer, Nick Dufour, Thomas Leung, Christoph Bregler, Stephanie CY Chan
Surfacing actionable insights for improved evals and models through a comprehensive analysis of LLM competency gaps.
Maty Bohacek, Ignacio Vilanova Echavarri
Leveraging media provenance to construct and evaluate compliant datasets for generative AI training.
Vaclav Knapp, Maty Bohacek
Shows synthetic human motion data improves action classification through pose transfer techniques.
Tomas Peterka, Maty Bohacek
Dataset of news articles with provenance metadata for assessing media relevance.
Saad Hassan, Maty Bohacek, Chaelin Kim, Denise Crochet
Exploring design and usage experience of an AI-driven video-based ASL dictionary with learners.
Tomas Peterka, Maty Bohacek
Using LLMs and provenance metadata to determine relevance of images and videos in news stories.
Vaclav Knapp, Maty Bohacek
Evaluating the realism of human motion generated by pose transfer models.
Maty Bohacek, Hany Farid
Method to determine whether generative AI models were trained on specific images using black-box inference.
Maty Bohacek, Hany Farid
Distinguishing real and fake (T2V) video using multi-modal embeddings, evaluated on the DeepAction dataset.
Maty Bohacek, Hany Farid
Demonstrates how generative models retrained on their outputs produce distorted images with lasting effects.
Vaclav Knapp, Maty Bohacek
LLM-based interface that provides pose-aware feedback to sign language learners.
Maty Bohacek
Evaluating methods for erasing artistic style concepts from diffusion models.
Maty Bohacek
Documenting the creation of an AI news anchor and its implications for the disinformation landscape.
Maty Bohacek, Michal Bravansky
Case study showing XGBoost can outperform GPT-4 on certain text classification tasks.
Maty Bohacek, Hany Farid
Analysis of the creation and implications of AI-generated news anchors for media integrity.
Maty Bohacek
Analyzing Slovakia's experience with deepfake-enabled election interference and lessons for other democracies.
Maty Bohacek, Hany Farid
Robust technique to detect lip-sync deepfakes by comparing audio-to-text and lip-reading transcriptions.
Maty Bohacek, Saad Hassan
Initial design and user evaluation of a video-based ASL dictionary system allowing real-time and pre-recorded sign lookup.
Maty Bohacek, Michal Bravansky, Filip Trhlik, Vaclav Moravec
A dataset for article-level trustworthiness assessment in Czech news media.
Maty Bohacek
Exploring the impact of large language models on education and academic integrity.
Jakub Kanis, Ivan Gruber, Zdenek Krnoul, Maty Bohacek, Jakub Straka, Marek Hruz
Multi-stage transformer architecture for hand pose estimation from full-scene depth images.
Maty Bohacek, Marek Hruz
Few-shot sign language recognition leveraging existing online dictionary resources.
Maty Bohacek, Hany Farid
Geometric and photometric analysis distinguishing GAN and diffusion model synthesized faces.
Maty Bohacek
Using news source classification as a proxy for detecting misinformation in non-article texts.
Maty Bohacek, Hany Farid
Biometric-based approach to protect world leaders against deepfakes using facial, gestural, and vocal mannerisms.
Maty Bohacek, Zhuo Cao, Marek Hruz
Achieves state-of-the-art on WLASL100 using MediaPipe pose estimation, being 11x faster than comparable methods.
Maty Bohacek, Hany Farid
Applying deepfake detection techniques to protect President Zelenskyy from synthetic media attacks.
Marek Hruz, Ivan Gruber, Jakub Kanis, Maty Bohacek, Miroslav Hlavac, Zdenek Krnoul
Ensemble methods for improving isolated sign language recognition accuracy.
Maty Bohacek, Marek Hruz
Novel transformer architecture for word-level sign language recognition using skeletal pose data.
Ivan Gruber, Zdenek Krnoul, Marek Hruz, Jakub Kanis, Maty Bohacek
Exploring how different data modalities support each other in sign language recognition tasks.
Maty Bohacek
Czech news article dataset and methods for source-level credibility assessment in low-resource contexts.