Prof. Dr. Vera Schmitt | XplaiNLP Research Group

Privacy for Smart Speech Technology (PSST)

Tue, 01 Apr 2025 00:00:00 +0000

Project Overview

The rapid growth of smart speech technology, much of which is operated by companies outside the European Union, has transformed how consumers access information and applications by using just their voices. Recently, however, we’ve learned that devices like smart speakers and smartwatches may expose users to privacy risks. Both service providers and hackers can capture not only what users say but also sensitive information hidden in their voices (like health information), all without the users knowing or agreeing to it.

To tackle these risks, new privacy-protecting smart speech technologies could help create business models that comply with EU laws like the GDPR and the AI Act. The PSST joint doctoral training network will train new researchers to develop these vital technologies using advanced deep learning methods. PSST will work on removing identifying features in speech, improving interactions between devices and cloud services, and creating new ways to assess privacy threats.

The PSST consortium includes top research labs in speech processing, privacy, and usability, as well as several innovative companies. This academic-industrial partnership will offer a robust training program that equips doctoral researchers with skills in speech processing, machine learning, and a thorough understanding of privacy issues and technologies. Together, PSST partners and researchers will not only offer a unique, user-focused perspective on smart speech technologies but also encourage innovative business models for responsible and sustainable privacy-aware technologies in the EU.

PSST is a joint doctoral training programme, a consortium of 7 European universities and 11 industrial partners, funded by Horizon Europe Marie Skłodowska-Curie Action, the European Union’s flagship funding programme for doctoral training.

The project is organized into three teams: Defence, Attack, and Utility. The teams will engage in a friendly competition for the iterative advancement of privacy-preserving speech technology. Each team has four doctoral researchers working on distinct projects.

Partners:

AALTO University
EUROCOM GIE
INESC ID
Stitching Radboud University
Institut National De Recherche
Ruhr-Universität Bochum
TU Berlin

news-polygraph

Mon, 01 May 2023 00:00:00 +0000

Project Overview

news-polygraph is a collaborative research project working on a comprehensive, multimodal technology platform for analyzing and detecting disinformation. By combining AI analysis methods for text, audio, image and video as well as the integration of crowd support, our project aims to support journalists and the media in effectively combating disinformation.

The news-polygraph research alliance is the biggest research project funded by the Federal Ministry of Research, Transfer and Space (BMFTR previously BMBF) as part of the “RUBIN – Regional Entrepreneurial Alliances for Innovation” funding program. A total of nine partners from Berlin, Brandenburg and Thuringia, including institutions from science, media, technology and business, are working together on effective solutions to integrate AI as intelligent decision support in the information verification workflow.

news-polygraph helps to strengthen trust in the media and information by supporting the integrity of journalistic work and protecting the public from misleading content.

FakeXplain

Fri, 01 Nov 2024 00:00:00 +0000

Project Overview

FakeXplain aims to develop transparent and meaningful explanations for AI-based systems that detect disinformation, addressing the challenge of making these explanations understandable for non-technical users like journalists and citizens. Given the rising influence of generative AI models in spreading misinformation, this project focuses on creating explanations that build trust and comply with the upcoming EU AI regulations.The project will investigate various explanation methods, such as attribution techniques and chain-of-thought prompting, while developing criteria to evaluate these explanations based on factors like transparency, robustness, and user trust.

Partners:

Fraunhofer HHI
DFKI
Q&U Lab Technische Universität Berlin
Tel Aviv University

VeraXtract

Sun, 01 Sep 2024 00:00:00 +0000

Project Overview

The VeraXtract project is funded by the BMFTR to establish and expand the XplaiNLP research group. The thematic focus of the project is on the extraction and detection of disinformation narratives and the development of approaches in the field of eXplainable AI.

Within the project, PhD students will focus on different topics:

Knowledge Base: development of a vector-based database consisting of disinformation narratives and provenance analysis/similarity search,
Narrative Extraction: development of LLM-based methods and approaches for disinformation narrative recognition and extraction
xAI: Development of AI-based explanations such as Chain-of-Thought-Prompting and Mechanistic Interpretability to communicate results to different user groups

Goals

In the long term, the knowledge and experience gained from VeraXtract should help to strengthen society’s resilience to disinformation. The solutions developed also have potential for use in content moderation in social media and could serve as the basis for new standards in the fight against disinformation in the future.

In VeraXtract, we want to create an overview and go into less depth. This overview is necessary to provide better information about disinformation and also to be able to more quickly assign statements to existing disinformation narratives. The results of the VeraXtract project will be integrated into a publicly accessible platform.

VERANDA

Mon, 01 May 2023 00:00:00 +0000

Project Overview

The primary aim of VERANDA is to explore and promote informational self-determination in the context of personal and sensitive medical data, especially among stigmatized groups. This includes advancing essential medical diagnostics, such as cancer detection and clinical-psychological care, which has seen increased demand during the COVID-19 pandemic. The project is closely tied to the provision of personal data for research purposes, highlighting the importance of effective risk-opportunity communication with patients and the public. By examining the conditions under which patients are willing to share their data during online consultations, the project seeks to develop communication strategies that enhance users’ understanding of relevant data protection technologies.

Partners:

TU Berlin (XplaiNLP)
Fraunhofer AISEC
BIH and IfSS Charité
DFKI (SLT)

ORCHESTRA

Sat, 20 Apr 2024 00:00:00 +0000

Extending Information Bottleneck Attribution to Video Sequences for Deepfake Detection

Sat, 18 Apr 2026 00:00:00 +0000

From Weights to Activations: Is Steering the Next Frontier of Adaptation?

Wed, 15 Apr 2026 00:00:00 +0000

Transparency as Architecture: Structural Compliance Gaps in EU AI Act Article 50 II

Fri, 27 Mar 2026 00:00:00 +0000

Gendered Prompting and LLM Code Review: How Gender Cues in the Prompt Shape Code Quality and Evaluation

Wed, 25 Mar 2026 00:00:00 +0000

Multiperspectivity as a Resource for Narrative Similarity Prediction

Mon, 23 Mar 2026 00:00:00 +0000

Retrieving Climate Change Disinformation by Narrative

Mon, 23 Mar 2026 00:00:00 +0000

News Credibility Assessment by LLMs and Humans: Implications for Political Bias

Sun, 01 Mar 2026 00:00:00 +0000

Selective Multimodal Retrieval for Automated Verification of Image–Text Claims

Sun, 01 Mar 2026 00:00:00 +0000

Take It All: Ensemble Retrieval for Multimodal Evidence Aggregation

Sun, 01 Mar 2026 00:00:00 +0000

Persona Prompting as a Lens on LLM Social Reasoning

Wed, 28 Jan 2026 00:00:00 +0000

Order in the Evaluation Court: A Critical Analysis of NLG Evaluation Trends

Mon, 12 Jan 2026 00:00:00 +0000

Can Large Language Models Still Explain Themselves? Investigating the Impact of Quantization on Self-Explanations

Thu, 01 Jan 2026 00:00:00 +0000

Parallel Universes, Parallel Languages: A Comprehensive Study on LLM-based Multilingual Counterfactual Example Generation

Thu, 01 Jan 2026 00:00:00 +0000

AI Development and Governance: Navigating Trust, Transparency, Innovation, and the Challenges of Information Warfare

Sat, 01 Nov 2025 00:00:00 +0000

Deepfakes–Unsere neue Realität?

Wed, 08 Oct 2025 00:00:00 +0000

Truth or twist? optimal model selection for reliable label flipping evaluation in llm-based counterfactuals

Wed, 01 Oct 2025 00:00:00 +0000

FitCF: A Framework for Automatic Feature Importance-guided Counterfactual Example Generation

Fri, 25 Jul 2025 00:00:00 +0000

EU-CSA FIMI-RESIST)

Tue, 20 May 2025 00:00:00 +0000

Associated Partners:

Gretchen.AI
Alexandra Geese

Utilising Large Language Models for Adversarial Attacks in Text-to-SQL: A Perpetrator and Victim Approach

Mon, 03 Mar 2025 00:00:00 +0000

Towards Automated Fact-Checking of Real-World Claims: Exploring Task Formulation and Assessment with LLMs

Thu, 13 Feb 2025 00:00:00 +0000

ITG Position Paper: Large Language Models are Transformers in Artificial Intelligence, Industry, Education, and Society

Wed, 27 Nov 2024 00:00:00 +0000

Anchored Alignment for Self-Explanations Enhancement

Thu, 17 Oct 2024 00:00:00 +0000

From Construction to Application: Advancing Argument Mining with the Large-Scale KIALOPRIME Dataset

Wed, 18 Sep 2024 00:00:00 +0000

Cross-Refine: Improving Natural Language Explanation Generation by Learning in Tandem

Wed, 11 Sep 2024 00:00:00 +0000

Disentangling User States in QoE: Situation-Dependent and Independent Factors

Tue, 18 Jun 2024 00:00:00 +0000

Evaluating Human-Centered AI Explanations: Introduction of an XAI Evaluation Framework for Fact-Checking

Mon, 10 Jun 2024 00:00:00 +0000

Implications of Regulations on Large Generative AI Models in the Super-Election Year and the Impact on Disinformation

Mon, 10 Jun 2024 00:00:00 +0000

NewsPolyML: Multi-lingual European News Fake Assessment Dataset

Mon, 10 Jun 2024 00:00:00 +0000

The Role of Explainability in Collaborative Human-AI Disinformation Detection

Mon, 03 Jun 2024 00:00:00 +0000

How Much is Your Instagram Data Worth? Economic Perspective of Privacy

Tue, 23 Apr 2024 00:00:00 +0000

Deutsch-Israleische Projektkooperation (DIP)

Sat, 20 Apr 2024 00:00:00 +0000

Associated Partners:

Prof. Ran Balicer Ben-Gurion University of the Negev
Prof. Marc Dewey Charité
Prof. Leif Sander Charité
Prof. Sebastian Möller TU Berlin

Fake-O-Meter

Sat, 20 Apr 2024 00:00:00 +0000

Partners:

DFKI (SLT)
TU Berlin (XplaiNLP Group)
Otto-von-Guericke-Universität Magdeburg Fachgebiet Mobile Dialogsysteme (MDS)
HS Magdeburg/Stendal Institut für demokratische Kultur
Deutsche Welle
Deutsche Presse-Agentur GmbH
Rundfunk Berlin-Brandenburg

KI-basiertes Kreislaufwirtschaft Ökosystem (KIKÖ)

Sat, 20 Apr 2024 00:00:00 +0000

Partners:

DFKI (SLT)
TU Berlin (XplaiNLP Group)
CircularTree GmbH

Classifying Sustainability Reports Using Companies Self-Assessments

Thu, 21 Mar 2024 00:00:00 +0000

Towards a Computational Framework for Distinguishing Critical and Conspiratorial Texts by Elaborating on the Context and Argumentation with LLMs

Wed, 31 Jan 2024 00:00:00 +0000

Augmented Political Leaning Detection: Leveraging Parliamentary Speeches for Classifying News Articles

Mon, 01 Jan 2024 00:00:00 +0000

Privacy-Aware Decision Making: The Effect of Privacy Nudges on Privacy Awareness and the Monetary Assessment of Personal Information.

Mon, 01 Jan 2024 00:00:00 +0000

What is Your Information Worth? A Systematic Analysis of the Endowment Effect of Different Data Types

Wed, 08 Nov 2023 00:00:00 +0000

How Risky is Multimodal Fake News Detection? A Review of Cross-Modal Learning Approaches under EU AI Act Constrains

Sat, 19 Aug 2023 00:00:00 +0000

A Transfer Learning Approach for SDGs Classification of Sustainability Reports

Fri, 21 Jul 2023 00:00:00 +0000

Comparing Simulated and Real Conversations for QoE Assessments: Insights from ARKit-Based Facial Configuration Analyses

Tue, 20 Jun 2023 00:00:00 +0000

Unraveling the Hangry Rater: Non-linear Effects of Hunger on Multimedia Quality Perception

Tue, 20 Jun 2023 00:00:00 +0000

ateSDG

Wed, 01 Mar 2023 00:00:00 +0000

Project Overview

The European Union’s 17 Sustainable Development Goals are already deeply embedded in business and society. However, it has been difficult to measure the contributions of the various actors. The ateSDG project is working to take a closer look at organizations from Berlin and Brandenburg by categorizing published texts using Natural Language Processing. In addition to classification, another focus of the project is model explanation. The selection and visualization of “explainability features” allows people to understand and validate the actions of the model. The goal of the project is an interactive dashboard that automatically assigns SDGs when text is entered and illustrates the underlying text modules and facts.

Team

Funding

By: Climate Change Center Berlin Brandenburg
Total Amount: EUR 60,000
Duration: March 2023 to December 2023

Fighting Disinformation: Overview of Recent AI-Based Collaborative Human-Computer Interaction for Intelligent Decision Support Systems

Sun, 19 Feb 2023 00:00:00 +0000

Multi-Task Learning for German Text Readability

Sun, 01 Jan 2023 00:00:00 +0000

A Feature Extraction based Model for Hate Speech Identification

Tue, 11 Jan 2022 00:00:00 +0000

Towards Automatic Green Claim Detection

Mon, 13 Dec 2021 00:00:00 +0000

Implications of the New Regulation Proposed by the European Commission on Automatic Content Moderation

Wed, 10 Nov 2021 00:00:00 +0000

Towards a Novel Benchmark for Automatic Generation of ClaimReview Markup

Mon, 21 Jun 2021 00:00:00 +0000

DFG project

Sat, 01 Feb 2020 00:00:00 +0000

Project Description

The aim of this project is to assess how much individuals would pay for protecting their personal data, such as location information in contrast to how much they would request when selling the information to different data requestors. The monetary assessment of privacy is analysed in relation to the individual’s privacy concern, privacy literacy and privacy awareness to understand influencing factors for the monetary assessment of personalised information.

Expected Outcomes

Assessment of existing methods for the monetary valuation of personal data
Better understanding of influencing and moderating factors, such as privacy concern, privacy literacy and personality traits
Analysis of the structural relation of the different influencing and moderating factors

Team

Vera Schmitt

Funding

By: Deutsche Forschungsgemeinschaft (DFG) - MO 1038/28-1
Total Amount: EUR 233,220
Duration: February 2020 to January 2023