AI4Media Technological Highlights - Discover AI4Media's Key Research Outcomes on Trustworthy AI

This booklet showcases the highlights of four years’ work towards research and development in Trustworthy AI. The relevant work package developed new techniques and algorithms for the application of trustworthy AI for media industry use cases, across four key pillars of interest: (i) adversarial robustness (covering novel attacks and defences for evasion and data poisoning, and certifying and verifying the robustness of AI models), (ii) explainability and interpretability (including assessing the explainability of multi-modal deep-fake detection classifiers), (iii) privacy and security (where privacy and security is preserved on the training data of AI models), and, (iv) fairness (concerning the detection and mitigation of bias in AI models across a variety of data modes and types of AI models).