Trustworthy AI – WP3

Explainability, Safety, Fairness, Accountability, Privacy, and Sustainability are the dimensions of Trustworthy AI that are necessarily intertwined with the foundation themes of the project through a continuous mutual exchange of requirements and challenges to develop legal protection and value-sensitive approaches. The questions that will drive the research are:

• How can we guarantee user trust in AI systems through explanation? How to formulate explanations as Machine-Human conversation depending on context and user expertise?
• How to bridge the gap from safety engineering, formal methods, verification as well as validation to the way AI systems are built, used, and reinforced?
• How can we build algorithms that respect fairness constraints by design through understanding causal influences among variables for dealing with bias-related issues?
• How to uncover accountability gaps w.r.t. the attribution of AI-related harming of humans?
• Can we guarantee privacy while preserving the desired utility functions?
• Is there any chance to reduce energy consumption for a more sustainable AI and how can AI contribute to solving some of the big sustainability challenges that face humanity today (e.g. climate change)?
• How to deal with properties and tensions of the interaction among multiple dimensions? For instance, accuracy vs. fairness, privacy vs. transparency, convenience vs. dignity, personalization vs. solidarity, efficiency vs. safety and sustainability.

Coordinated Actions and tasks

Coordination Actions (CA) are groups of researchers convened around a specific topic of interest, to start, to investigate, to promote and to accelerate Trustworthy AI. Interested to join? Please take a look at the current CA proposals. You may join to some existing one, just contact the CA leader, or propose one by yourself. To do so, contact the WP3 Task Leader of the task that is predominant in your proposal.

T3.1 Explainable AI Systems

Comparison of methods for interpretation of convolutional neural networks for the classification of multiparametric MRI images on unbalanced datasets. Case study: prostate cancer, vestibular schwannoma cancer

Tasks: T3.1 Explainability , T3.7 Trustworthy AI as a whole, T4.3: Learning and reasoning with embeddings, knowledge graphs & ontologies

Partners: CNR, UNIPI, UGA, INRIA, LIRA

Explainable malware/security threat detection: Comparison of methods for detection and prediction of malware/security attacks that are able to produce some kind of explanation or characterization of the attack

Tasks: T3.1 Explainability, T3.7 Trustworthy AI as a whole, T4.3: Learning and reasoning with embeddings, knowledge graphs & ontologies

Partners: Slovak.AI, CNR

T3.2 Safety and Robustness

Dealing with truly adversarial examples

Tasks: T3.2 Safety and Robustness , T3.4 Accountability and Reproducibility

Partners: VRAIN/UPV, JRC-EC/VRAIN,, Slovak.AI, CNR

Robust Evaluation: prevent specialisation and test replacement

Tasks: T3.2 Safety and Robustness , T3.4 Accountability and Reproducibility

Partners: VRAIN/UPV, JRC-EC/VRAIN, CEA

SafeAI and AISafety workshops (AAAI and IJCAI)

Tasks: T3.2 Safety and Robustness , T3.8 Fostering the AI scientific community around Trustworthy AI

Partners: CEA, VRAIN/UPV, UNIPI, CNR, TUDelft

Formal methods and V&V for AI

Tasks: T3.1 Explainability, T3.2 Safety and Robustness , T3.4 Accountability and Reproducibility, T3.7 Trustworthy AI as a whole, T3.8 Fostering the AI scientific community around Trustworthy AI

Partners: CNR, VRAIN/UPV

T3.3 Fairness, Equity, and Justice by Design

Operationalizing Fairness Metrics

Tasks: T3.3 Fairness, Equity, and Justice by Design

Partners: UNIPI, INRIA, TUDelft

T3.4 Accountability and Reproducibility

Emergent responsibility in reproducible multi-agent settings

Tasks: T3.4 Accountability and Reproducibility, T6.4 Emergent Behaviour, agent societies and social networks

Partners: TUDelft

Holistic assessment and certifications of AI systems for reproducibility and accountability

Tasks: T3.2 Safety and Robustness , T3.4 Accountability and Reproducibility, T3.7 Trustworthy AI as a whole

Partners: CNR, TUDelft, PUT, TUE

T3.5 Respect for Privacy

Impact on Fairness of Privacy Preserving Data Transformation

Tasks: T3.3 Fairness, Equity, and Justice by Design, T3.5 Respect for Privacy

Partners: UNIPI, UGA, LiU, Inria

Challenges for Guaranteeing Privacy while Preserving Utility

Tasks: T3.5 Respect for Privacy, T3.7 Trustworthy AI as a whole

Partners: DFKI, EPFL, INRIA, LIU

Automatic Tools for Analyzing and Explaining Privacy Risks

Tasks: T3.1 Explainability, T3.5 Respect for Privacy, T3.7 Trustworthy AI as a whole

Partners: UGA, CNR, UNIPI, INRIA

T3.6 Sustainability

Probabilistic Workload Forecasting in Cloud Computing using Deep Learning approaches

Task: T3.6 Sustainability

Partners: UCC