Metamorphic Testing for Trustworthy AI
Srinivas Padmanabhuni, Neelima Vobugari · 2024
Artificial intelligence (AI) has penetrated our daily lives so much that it is impossible to imagine life without AI. However, there has been a spate of accidents involving AI systems, leading to serious losses including loss of lives, financial fines, and loss of reputation for the relevant AI providers. All these accidents typically lead to a lack of trust in AI, resulting in gaps in the implementation of trustworthy AI. The primary root cause of these accidents is the lack of sufficient testing of AI systems, by the relevant developers and quality professionals involved in the development of AI solutions. In this chapter, we highlight the challenges involved in the testing of AI solutions, primarily those powered by Machine Learning (ML), due to the lack of a test oracle. We then present a solution to the lack of a test oracle in AI systems by using a metamorphic testing-based approach to test AI systems. We detail the different metamorphic testing-based approaches to testing AI systems. We hope this chapter can present viable approaches for AI implementers and testers to adopt to guarantee trustworthy AI, minimizing accidents.