Titulo Estágio
AuditAI - Co-designing web platform for external algorithm auditing
Local do Estágio
Porto or Remote Work
Enquadramento
Background and motivation
Artificial intelligence is more and more common. Today, we see algorithms supporting diagnosis in healthcare, driving in autonomous cars, or determining jail time in court decisions. Drawing on algorithms for making decisions has the potential to make decisions fairer, as the same criteria is always applied, but if algorithms are developed with biased datasets or wrong assumptions, they can increase discrimination or inappropriate outcomes.
To investigate the appropriateness of algorithms we rely on external auditing. By gaining access to algorithms, researchers are able to experiment different conditions and evaluate their algorithm outputs. However, in many areas algorithms are close-sourced and difficult to access, in an effort to protect intellectual property that took years to create. The lack of access to the source or the algorithm running makes it extremely hard to audit algorithms and reduces the scrutiny and oversight of society over algorithms and the companies that use them.
Objetivo
Objectives
This thesis will create a platform for externally auditing algorithms. Drawing on in-depth interviews with researchers and developers creating close-sourced algorithms for different areas, the candidate will understand the issues of sharing code and datasets. Moreover, drawing on in-depth interviews with researchers auditing algorithms, the candidate will investigate the requirements for performing external audits. Based on these insights, the candidate will create a web platform where researchers and developers can deposit an executable of their algorithm and have other researchers externally audit it.
Innovative aspects
Platforms for externally auditing close-sourced algorithms are non-existent or rare. Related work consists of tools for testing artificial intelligence algorithms, e.g. RapidMiner or WEKA, where users can experiment different algorithms with the same dataset. Software testing tools, e.g. Jenkins, could also be considered related, but their role is to support developers in testing algorithms they are creating, and not to enable researchers to audit others’ algorithms.
Plano de Trabalhos - Semestre 1
Workplan
1 - Literature review and competitive analysis of systems for externally auditing black box algorithms;
2 - In-depth interviews with researchers and developers creating algorithms;
3 - Development of a fully-fledged project proposal.
Plano de Trabalhos - Semestre 2
Workplan
1 - In-depth interviews with researchers and developers creating algorithms;
2 - In-depth interviews with researchers who audit algorithms;
3 - Prototyping and/or co-design of web platform for auditing algorithms;
4 - Initial evaluation of the suggested concept.
Condições
Candidate Profile
- Good social skills for interacting with participants;
- Interest in understanding issues of developers and researchers creating/auditing algorithms;
- Willingness to perform interdisciplinary work;
- Proficiency in English.
Observações
MSc Thesis to be supervised by Prof. Paula Alexandra Silva (Department of Computer Science).
References
- Sandvig, C., Hamilton, K., Karahalios, K., & Langbort, C. (2014). Auditing Algorithms: Research Methods for Detecting Discrimination on Internet Platforms. Paper presented to “Data and Discrimination,” a pre-conference of the 64th annual meeting of the International Communication Association, Seattle, WA, USA.
- Crawford, K., Calo, R. (2016) ‘There is a blind spot in AI research’, Nature, October 13.
- https://ainowinstitute.org/aap-toolkit.pdf
Orientador
Francisco Maria Cruz Nunes
francisco.nunes@fraunhofer.pt 📩