TRIZ-Informed Validation of Predictive AI in a Holistic Cybersecurity Sandbox: A Machine Learning Pipeline for Trustworthy Forecasting
We propose a TRIZ-Informed Machine Learning (TIML) framework as part of an integrated cybersecurity sandbox to examine and validate AI-driven decision support systems. The framework employs six of the TRIZ creative principles as a sophisticated AI vulnerability diagnostic protocol. This framework investigates and addresses several specific adversarial concerns: COVID-19-induced structural breaks and cross-continuous entity scale heterogeneity. The sandbox validation approach improves random forest validation R² from 0.7251 to 0.8978, and MAPE improves from 11.99% to 6.83%. This work develops a method for a standards-oriented validation and evaluation framework within the bounds of the EU’s AI Act and combines the principles of causal irreversibility and temporal autocorrelation. The TIML framework is introduced as a modular sandbox testing framework for use with other predictive AI systems that are operating in adversarial conditions and are seeking real-world validation.
Authors
- Monica Leba (ORCID: https://orcid.org/0000-0002-4083-9121)
- Ionela Samuil (ORCID: https://orcid.org/0000-0002-5579-5598)
- Andreea Cristina Ionica (ORCID: https://orcid.org/0000-0003-1988-9340)
- Nicusor Gelu Constantin
- Isabella Fodor
- Eugen Florian Popescu
- Dragos Dabija
Institutions
- Universitatea Din Petrosani (RO)
Publication Details
- Journal
- International Journal of Education and Information Technologies
- Published
- 2026-11-18
- DOI
- https://doi.org/10.46300/9109.2027.21.2
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- article
- Field-Weighted Citation Impact
- 0.00