WP1: WP1 made various contributions to advancing AI security research, particularly regarding applications in the cybersecurity and autonomous driving use cases. Tools for the automated robustness assessment of malware detectors were developed and novel certification mechanisms against evasion and poisoning attacks were explored. The robustness of large language models (LLMs) and multimodal models against various threats was investigated. WP1 introduced new approaches to improving model efficiency through neural network pruning and model editing. Pruning techniques for attack mitigation were categorized and evaluated. The relationship between the accuracy and trustworthiness of AI models were analyzed, proposing a unified framework and technical and ethical metrics. Moreover, WP1 extended the evaluation of AI-generated image detectors with an adversarial robustness perspective, in addition to interpretability.
WP2: Jointly developed challenges for Document Intelligence and Health have been featured in leading conferences in the fields. The joint work with use-case researchers have established privacy into Document Intelligence research field. In privacy research, the main outreach event was a one-week workshop in Italy with 10+ leading researchers as invited speakers and 40+ participants. The discussions served as foundation to two reports currently being developed that are expected to shape the research field into the future. Research has progressed as planned and made significant contributions to private and collaborative learning. Changes in the broader research environment have led us to focus significantly on privacy with foundation models.
WP3: Significant progress was made in developing interpretable and trustworthy AI systems through a range of prototype-based, explainable-by-design deep learning methods. These were tested in various use cases. Key implemented software tools were evaluated for technical maturity and practical applicability under real-world conditions. Legal and ethical aspects of AI g oversight were addressed through interdisciplinary research and stakeholder engagement, including policy briefs, workshops, and interviews. These efforts improved coordination with industry partners and contributed to shaping the discourse on AI safety in line with the EU AI Act. WP3's work on human agency and oversight continued to identify and define challenges related to AI certification and use through workshops and interviews on the use of AI in various institutions.
WP4: The Use Case benchmarks were the basis for implementing two courses of action planned within the project. During this review period, 11 competitions on the different Use Cases through the ELSA Benchmarks Platform were hosted and linked to events relevant to the target scientific communities through workshops or tutorials. A complementary course of action within the ELSA consortium aimed to create new solutions for the real-life challenges that Use Cases put forward, based on the research outcomes of the project. During the reporting period, the collaboration between WP 1, 2 ,3 and 4 was strengthened, leading to baseline methods created by the consortium for the benchmarks set out. A WP4 workshop was organized to define the framework for benchmarks and specifications of the 6 use cases, aligning with the research plans of WP1, 2 and 3. An established benchmark platform enables data downloads, results submission and visualization, thus providing access to all use case resources and participation in ELSA competitions.
WP5 and 6: ELSA aims to be a catalyst for research on safe and secure AI in Europe, disseminating its research results to the scientific community, industry, and the general public. ELSA’s first Industry Call was concluded while the second Industry Call was launched. In the granted one-year extension of the project, the consortium will address current developments in the field of AI, particularly EU AI Act, and resolve specific issues. The dynamics of the expanded ELSA network and collaboration with industry partners and other relevant stakeholders can be intensified. The collaborating with strategic research partners will be continued.
