首页|Quantitative assessment of machine learning reliability and resilience

Quantitative assessment of machine learning reliability and resilience

扫码查看
原文链接
  • NETL
  • NSTL
  • Wiley
Advances in machine learning (ML) have led to applications in safety-critical domains,including security, defense, and healthcare. These ML models are confronted withdynamically changing and actively hostile conditions characteristic of real-world applications,requiring systems incorporating ML to be reliable and resilient. Many studiespropose techniques to improve the robustness of ML algorithms. However, fewerconsider quantitative techniques to assess changes in the reliability and resilience ofthese systems over time. To address this gap, this study demonstrates how to collectrelevant data during the training and testing of ML suitable for the applicationof software reliability, with and without covariates, and resilience models and the subsequentinterpretation of these analyses. The proposed approach promotes quantitativerisk assessment of ML technologies, providing the ability to track and predict degradationand improvement in the ML model performance and assisting ML and systemengineers with an objective approach to compare the relative effectiveness of alternativetraining and testing methods. The approach is illustrated in the context of an imagerecognition model, which is subjected to two generative adversarial attacks and theniteratively retrained to improve the system’s performance. Our results indicate thatsoftware reliability models incorporating covariates characterized the misclassificationdiscovery process more accurately than models without covariates. Moreover, theresilience model based on multiple linear regression incorporating interactions betweencovariates tracks and predicts degradation and recovery of performance best. Thus, softwarereliability and resilience models offer rigorous quantitative assurance methods forML-enabled systems and processes.

adversarial attacksmachine learningresiliencesoftware reliability

Zakaria Faddi、Karen da Mata、Priscila Silva、Vidhyashree Nagaraju、Susmita Ghosh、Gokhan Kul、Lance Fiondella

展开 >

Department of Electrical and ComputerEngineering, University of MassachusettsDartmouth, Dartmouth, Massachusetts, USA

Department of Computer Science, StonehillCollege, Easton, Massachusetts, USA

Department of Computer Science andEngineering, Jadavpur University, Kolkata, India

Department of Computer and InformationScience, University of Massachusetts Dartmouth,Dartmouth, Massachusetts, USA

展开 >

2025

10.1111/risa.14666

Risk analysis

Risk analysis

ISSN：0272-4332
年,卷(期)：2025.45(4)
  • 80