Trustworthy AI: Accuracy, explainability and interpretability, privacy, reliability, robustness, safety, and security or resilience to attacks

To use AI in human-critical applications that could affect well-being, environment, lifestyle, etc., we need to trust the AI systems and those who develop and implement them. The National AI Initiative notes:

“To be trustworthy, AI technologies must appropriately reflect characteristics such as accuracy, explainability and interpretability, privacy, reliability, robustness, safety, and security or resilience to attacks – and ensure that bias is mitigated.”

This vertical will focus on these characteristics, highlighting their technical and business challenges. The following definitions have been adapted from many sources [1] [2] [3].

– Explainability and Interpretability. Understanding is a requirement for developing trust. Explanations are necessary to enhance understanding, trust, and informed decision making. Explainability is the extent to which the internal mechanics of a model can be explained in human terms. Interpretability is about the extent to which a cause and effect can be observed within a system.

– Transparency. The AI data, system, and business models must be transparent. Humans must be aware when they are interacting with an AI system, and must be informed of its capabilities and limitations. Ultimately, transparency improves traceability and accountability.

– Privacy, Governance Risk, and Compliance. Besides providing privacy and data protection, data governance mechanisms must also be used, to guarantee data quality and integrity, while ensuring legitimized access to data.

– Robustness, Security/Resilience to attacks. ML algorithms could be vulnerable to novel adversarial inputs. Trustworthy AI systems must have Robustness (the ability to withstand the effects of adversaries), including adversarial threats and attacks.

– Fairness/Bias. Hidden biases could lead to discrimination and exclusion of underrepresented/vulnerable groups. To prevent this, AI systems must include proper safeguards against bias and discrimination. Except for when the bias is intrinsic to the ML algorithm (Algorithm Bias), biases are usually introduced in the ML model through its training data set: Sample Bias, Prejudice Bias, Measurement Bias, and Exclusion Bias.

[1] https://www.onespan.com/blog/trustworthy-ai-why-we-need-it-and-how-achieve-it
[2] https://www.technologyreview.com/2020/03/25/950291/trustworthy-ai-is-a-framework-to-help-manage-unique-risk/
[3] https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai