[en] Artificial Intelligence (AI) systems are proliferating in our society due to their capacity to
simulate human intelligence, behaviors, and processes. The increased utilization of AI systems in society, especially in high-risk settings such as autonomous systems and healthcare, has been accompanied by an increased concern about the impact of AI systems on society. In recent years, vulnerabilities to algorithmic bias, adversarial attacks, and data breaches have resulted in the critical assessment of how AI systems can be designed to be inherently trustworthy.
This dissertation presents the key concepts of trustworthiness in AI systems, with a focus on identifying the challenges associated with designing, developing, and deploying collaborative AI. Towards this purpose, key elements of trustworthy AI are identified, culminating in a set of concise guidelines that developers can leverage in the development of trustworthy AI. Further, this dissertation explores how techniques initially created solely for privacy, specifically federated learning, can be leveraged to build trust in machine-learning environments.
Federated learning is assessed for its implications on trustworthy principles, with a particular focus on how privacy is established to enable collaboration between participants without the sharing of private data. The security of federated learning is further assessed by demonstrating the impact of targeted model poisoning attacks and an assessment of Byzantine-tolerant defense mechanisms to prevent and defend against such attacks. Further, the potential for federated learning to be leveraged for compliance with regulatory requirements is assessed.