Trabalho em anais de eventoOpen Access

Bridging the Robustness Gap: Stress-Testing Artificial Intelligence Algorithms

OLIVEIRA, S. B.; Lima Filho, E. B.; CORDEIRO, L. C

2026 • IEEE 2nd International Conference on Consumer Technology

Resumo

The increasing deployment of artificial intelligence (AI) elements in safety-critical systems exposes a major concern: models that prove reliable on clean data can be brittle under real-world perturbations. Besides, that may vary according to different model classes, e.g., supervised and unsupervised ones, which should guide practical applications. In that sense, numerous evaluation techniques do exist, but there is no structured scheme or results that allow for comparing and choosing between different approaches. As a response, we have conducted a study based on a methodology that applies controlled, synthetic perturbations to data, labels, and model parameters, focusing on the experimental evaluation of supervised and unsupervised algorithms. We have stress-tested 10 algorithms across 5 diverse datasets, whose outcome revealed a pronounced dichotomy: unsupervised models may suffer near-structural collapse, whereas supervised models can retain partial functionality due to the anchoring effect of labels. In conclusion, these results suggest systematic stress testing as a prerequisite for reliable and safer AI systems.

palavras-chave

Palavras-chave

robustness gapstress testingsafety-critical systemsdata perturbationparameter perturbationsupervised learningunsupervised learningstructural collapserobustness analysis

Status

Published

links