What is ELSA?
European Lighthouse on Secure and Safe AI (ELSA) is a growing network of excellence that spearheads efforts in foundational safe and secure AI methodology research. ELSA builds on and extends the internationally recognized and excellently positioned ELLIS (European Laboratory for Learning and Intelligent Systems) network of excellence.
ELSA is based on the understanding that safe and secure AI can only be achieved by building on foundational research. Relying on heuristics alone is susceptible to arms races and will not yield appreciable, reliable or sustainable levels of safety or security and is prone to fail overcoming the key obstacles in shaping AI in compliance with our European values. ELSA will overcome this by combining rigorous approaches such as robustness certification and differential privacy with machine learning and deep learning that are ubiquitous in applications. Hence, ELSA is based on a research agenda that encompasses the key research directions and fosters vital research by bringing together researchers across communities as well as different stakeholders. The research community is unified and directed towards addressing grand challenges that reach across use-cases by benchmarks that will drive innovation and provide measurable progress. The use-cases call for different, domain-specific requirements and risk analysis that require solutions, tools, and software that overcome obstacles in a trustworthy approach that conforms with legal and ethical principles grounded on European values. Our investigations emphasize state of the art methodology and technology (with a particular focus on Deep Learning techniques) in order to ensure practical relevance of the solutions.
Research Programmes
Building on the established ELISE framework for trusted AI, ELSA advances these foundations with a dedicated focus on safe and secure AI methodology. Shaped through continuous stakeholder collaboration and real-world use cases, the research agenda directly targets key technical obstacles in AI deployment.
ELSA comprises the following four research programmes, that interact with the ELLIS research programmes:
- Technical Robustness and Safety
- Privacy-Preserving Techniques and Infrastructure
- Human Agency and Oversight and
- ELSA Innovation Lab
The ELSA network
ELSA's founding members include European experts in all aspects of safe and secure AI, including technical robustness, privacy preserving techniques and human agency and oversight. In addition, ELSA brings on board research and industry experts in six different sectors that are key application areas of safe and secure AI.
It comprises 21 ELLIS units and 47 network partners all over Europe.
ELSA Benchmarks
The ELSA Benchmarks Platform provides a single access point to all ELSA Use Case related resources (data, metrics, benchmarks) and allows participation to the ELSA competitions.
Please browse the use cases, and find out about related events (competitions, workshops, tutorials) running. Each use case will host one or more scientific competitions over the duration of the ELSA project. Register in the platform in order to get access to download materials and to submit your own methods to the competitions.
Use Cases
ELSA develops six flagship use cases, each co-led by an academic and an industry partner, to test and demonstrate safe and secure AI methods in real-world settings.
Health - Federated Genome Medicine (EMBL, NVIDIA): privacy-preserving, federated analysis of genomic and clinical data across Europe.
Autonomous Driving - Robust Perception (MPG, Valeo): provably robust perception systems for self-driving vehicles under unexpected conditions.
Robotics - Learning through Human Interaction (KTH, PAL): robots that learn safely from human demonstration while protecting personal data.
Multimedia - Tackling Disinformation (UNIMORE, Leonardo): detecting deepfakes and manipulated media through combined technical and human-in-the-loop analysis.
Cybersecurity - Malware Detection (CINI, Pluribus One): AI-based malware detection systems that adapt reliably to new and evolving threats.
Document Intelligence - Document Visual Question Answering (CVC-CERCA, Yooz): privacy-preserving methods for extracting and reasoning over information in sensitive documents.
Explore all use cases at elsa-ai.eu/use-cases.
Research Agenda and Policy Context
A European Approach to Trustworthy AI
ELSA's research responds directly to the European vision for trustworthy AI and to requirements emerging from the EU AI Act. The aim is to ground safety and security guarantees in rigorous science rather than heuristics alone, so that AI systems deployed in Europe remain reliable, transparent and aligned with European values.
Three Grand Challenges
ELSA's Strategic Research Agenda organises its work around three Grand Challenges that complement the research programmes above. Technical Robustness and Safety aims to build AI systems with provable guarantees against adversarial and out-of-distribution inputs. Robust Private Collaborative Learning develops distributed learning methods that protect data privacy while remaining reliable. Human Agency and Oversight ensures AI systems stay interpretable, controllable and accountable to the people affected by them.
Case Study: Securing Large Language Models
As chatbots and other assistants built on large language models become widespread, ELSA is examining the safety challenges they raise, including protecting personal data seen during training, keeping outputs aligned with legal duties such as the GDPR, and keeping a human able to understand and evaluate what these systems produce.
Read the Strategic Research Agenda
For full details behind these challenges, mitigations and open research directions, see ELSA's Strategic Research Agenda (PDF).
