TRUMAN · TRUstworthy huMAN-centric artificial intelligence

Artificial Intelligence (AI) has become the main driver of growth in information technology, touching all sectors in the industry, ranging from healthcare, to finance, and transport. The goal of TRUMAN is to design and develop generic technologies and methodologies for improving AI systems’ resilience against security, privacy, and fairness attacks, as well as to increase the trust that their users have in these systems, while accounting for different phases of the AI life cycle, starting from data collection through training and deployment. TRUMAN will encompass three major AI architectures: knowledge graph (KG) representation and learning, continual learning, and large language models (LLMs). These three approaches will be investigated with the goal of designing solutions aimed for scenarios involving dynamic data collection, distributed model training, and scenarios that involve human-in-the-loop (HITL) design principles. TRUMAN will develop customized robustness solutions for both existing and newly developed privacy, adversarial, and fairness attacks. The project will also consider the impact of these solutions on humans, how to explain to them the underlying technologies and the risks, and employ their help in the improvement of these models. These technologies will be evaluated against four use cases illustrating scenarios from different sectors, i.e., marketing, IT, finance, and healthcare. Moreover, the TRUMAN project will strive to generate generic/holistic guidelines and recommendations for building trustworthy and robust AI systems.