Title: AI Privacy and Fairness Assessments

Duration: ongoing since April 2022

Research Area: Responsible AI

The adoption of Artificial Intelligence (AI) into digital services, e.g., in recruitment, medical applications or Smart-Home products, raises new challenges regarding ethics and privacy. AI algorithms can inherit biases from the training data, resulting in unfairly favoring certain demographics, leading to discrimination against others, providing inaccurate outcomes or decisions, etc. Furthermore, AI often depends on processing of vast amounts of personal data, and the widespread adoption of AI makes it the new normal to use such data without hesitation. In addition, AI algorithms typically operate as black boxes, i.e., the data included in the model and the decision-making is not transparent. This project uses interdisciplinary research in the research areas of computer science, ethics, and jurisprudence to strive for privacy-aware, fair AI applications. 

Aims

In this project, we follow three lines of research. First, we explore approaches such as Federated Learning, which allow training models without having to aggregate personal data at a centralized party, and without degrading the model quality, as is the case with other privacy-enhancing approaches.  Second, we strive to develop a tool chain, that allows us to evaluate fairness aspects such as biases, ethics, morale, etc. by using AI models and well-researched linguistic approaches.  Third, we want to use this tool chain to automatically assess fairness in privacy policies and other legal texts that have a high social relevance for users of AI-based services.

To this end, we identify various fairness dimensions that are relevant for AI, e.g., informational fairness, representational fairness and ethics/morality. We explore how these dimensions might appear in privacy policies. Finally, we strive for a qualitative and quantitative assessment of fairness aspects in a large corpus of texts. This is to provide authorities, companies, and jurists with evidence on best practices, an option to quickly identify outliers in a large text corpus, and with a tool to make recommendations. We also want to use this approach to identify fairness issues in the AI model itself, by letting it assess texts with well-known fairness problems as a reference.

Problem

To the best of our knowledge, we are the first to use AI to identify fairness issues in privacy policies. We seek a close collaboration with jurists and ethicists to develop holistic concepts that ensure a fair, transparent and privacy-aware use of AI approaches. Thus, our challenges are not only in the technical aspects of assessing fairness with existing approaches from large language models to linguistic approaches. We are also working on a common understanding of fairness issues from the perspectives of a jurist, ethicist, and computer scientist.

Technology

We make use of various large language models, natural language processing approaches, linguistic measures and federated learning schemes.

Outlook

At this stage of our project, we have a good understanding of promising lines of future research. We are currently exploring options to write an interdisciplinary research proposal, with the objective to taking our collaboration to the next level. We also plan to broaden our research towards federated learning and similar approaches, that allow to build high-quality models without having to transfer sensitive data.

Publications

• FREIBERGER, Vincent; BUCHMANN, Erik. Beyond the Fine Print: Fairness in Privacy Policies. (Accepted Poster) In: AI Ethics & Human-Computer Interaction Conference, Graz, 2024
  
• FREIBERGER, Vincent; BUCHMANN, Erik. Fairness Certification for Natural Language Processing and Large Language Models. (To Appear) In: Proceedings of the 10th Intelligent Systems Conference (IntelliSys’24), 2024
  
• FREIBERGER, Vincent; BUCHMANN, Erik. Fairness Certification for Natural Language Processing and Large Language Models. In: arXiv.2401.01262, 2024
  
• BARTELT, Bianca; BUCHMANN, Erik. Transparency in Privacy Policies. (To Appear) In: Proceedings of the 12th International Conference on Building and Exploring Web Based Environments (WEB’24), 2024
  
• HANNEMANN, Anika; ÜNAL, Ali B.; SWAMINATHAN, Arjhun; BUCHMANN, Erik; AKGÜN, Mete. A Privacy-Preserving Framework for Collaborative Machine Learning with Kernel methods. In: Proceedings of the 5th IEEE International Conference on Trust, Privacy and Security in Intelligent Systems and Applications (TPS-ISA’23), 2023
  
• HANNEMANN, Anika; ÜNAL, Ali B.; SWAMINATHAN, Arjhun; BUCHMANN, Erik; AKGÜN, Mete. A Privacy-Preserving Federated Learning Approach for Kernel Methods. In: arXiv:2306.02677, 2023

Team

Lead

• Prof. Dr. Erik Buchmann

Team Members

• Anika Hannemann
• Victor Jüttner
• Vincent Freiberger

Partners

• Prof. Dr. Birte Platow
• Dr. Hermann Diebel-Fischer
• Prof. Dr. Anne Lauber-Roensberg
• Julia Moeller-Klapperich
• Bianca Bartelt