Explainability Protocol and Methods (v2)


Deliverable D4.1 – 'Explainability protocol and methods' – reports the consortium work towards the development of explainable techniques for face recognition during the project. The deliverable has been published in two versions, one in month 12 and one in month 24. The first version (v1) already summarised relevant existing explainability techniques for AI face recognition solutions addressing both the cases of creating transparent models and post-hoc explanation of black-box models, described some plans and evaluation performance protocols, and provided a first analysis of ethical and legal issues. This is the second version of the deliverable which, in addition to v1 provides more detailed descriptions of the methods initially described in version 1, collects descriptions of additional methods developed and includes performance results obtained during experiments with their first implementations. The chapter addressing ethical and legal issues is extended and provides a more in depth look at various examinations of 'interpretability'-requirements according to the AI-Act proposal.