AMNESIA

In AMNESIA the algorithmic unfairness assessment problem will be faced by:

• Building a complete picture of the state-of-the-art of fairness metrics;
• Selecting, generalising, extending, and improving the identified metrics;
• Suggest ways to mitigate the unfair behaviour of the AI-based tools and models;
• Developing a fairness assessment Proof-of-Concept (PoC) and validating it using available
• FAAMG tools and public domain datasets;
• Assessing the commercial viability of the proposed solutions.

Specifically, AMNESIA will first review the notions of fairness and the families of methods able to impose these notions by listing their advantages and disadvantages. Then, new fairness notions able to generalize and extend the current ones to a wider framework will be proposed: in fact, most of the approaches are not able to deal with different problem categories like, for example, Machine Learning supervised and unsupervised problems or different data types like, for example, categorical and continuous sensitive features. Moreover, AMNESIA will develop a methodology and related tool at TRL 3 able to both semi-automatically select and exploit the best metrics for assessing the fairness of the specific AI-based FIT and to suggest actions to mitigate the unfair behaviour. Finally, AMNESIA will assess the commercial viability of the proposed approach. AMNESIA is coordinated by ZenaByte that is also in charge of developing the AMNESIA approach.ettings.