Audit AI- Co-designing Web Platform for External Algorithm Auditing

• Autor(a) principal: Fernandes, Rui Marques
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/138796
• topic: Auditing; Artificial Intelligence; Machine Learning; Algorithms; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Artificial Intelligence (AI) algorithms are becoming more crucial and decisive every day as the promise that these systems are able to mitigate human biases intensifies, thus pro- viding fair opportunities to everyone. When unchecked, algorithms might be subjected to bias and discriminatory outcomes, principally due to the use of biased datasets or even by defective implementations of the algorithms. Notwithstanding, there has not been a follow-up regarding the monitorization of the algorithmic systems, despite the increas- ing tendency in the use of these systems during decision-making processes. In order to increase the trustworthiness of these algorithms, auditing practices are starting to gain some relevance among companies, researchers, and developers. Nonetheless, the amount of auditing practices performed is still nowhere close the amount needed. Algorithmic auditing practices can prevent abuses by unethical stakeholders, provid- ing greater scrutiny to the algorithms by searching for discrimination and bias. However, ensuring that AI algorithms reflect societal values and do not present any severe case of bias and discrimination is not easy nor simple, considering that most of these algorithms are confidential and not easily accessible. The owners of the algorithms choose not to share their algorithm’s data to protect their intellectual propriety, which often took years to create. This lack of access to the algorithms’ source code, datasets, and results does not allow the auditor to inspect the algorithm entirely, thus making the auditing task far more difficult. This work aims to support the auditing of algorithms without accessing their source code by proposing a solution based on a web platform. The goal is to allow developers to connect their algorithms with the platform, which the auditors can then test using their datasets. So, it is possible to maintain the ”invisibility” of the algorithms’ source code, but at the same time, allow them to be audited. The platform was built based on interviews with developers, auditors, and researchers working in this space, which allowed us to understand and design an adequate solution that fulfilled the user’s requisites.