Trabalho em anais de evento
Modular Architecture for Robustness Assessment in AI Using Smart Inference
OLIVEIRA, S. B.; Lima Filho, E. B.; CORDEIRO, L. C
Resumo
Resumo
This article proposes a modular architecture to assess the robustness of artificial intelligence (AI) algorithms in synthetic inference scenarios. The approach introduces the smart inference pipeline (SIP), which applies configurable transformations to data, labels, and hyperparameters, simulating realistic noise and configuration errors. Five algorithms and five datasets were tested under controlled experiments, whose results show that perturbations consistently reduce model performance, with Gaussian naive bayes demonstrating the highest resilience. SIP supports reproducible and extensible experimentation, offering a solid foundation for robustness testing in AI pipelines.
palavras-chave
Palavras-chave
AI robustnessrobustness analysisnoise injectionconfiguration errorsGaussian naive Bayesmodel resiliencemodular architecture
Status
Published
links
Links e arquivos
Trabalho em anais de evento