Overview
SIST-TP EN ISO/IEC/TR 24029-1:2024 - Artificial Intelligence (AI) - Assessment of the Robustness of Neural Networks - Part 1: Overview (ISO/IEC TR 24029-1:2021) is a technical report that provides a comprehensive background on methods for assessing the robustness of neural networks in AI systems. Prepared by CEN-CENELEC Technical Committee JTC 21, this document highlights the importance of robustness in AI applications and describes the principal methodologies for evaluating how well neural networks maintain performance under diverse and challenging conditions.
With the increasing adoption of AI in areas such as industry, healthcare, transportation, and public services, ensuring AI system robustness is critical for reliability, safety, and trust. This report serves as an essential resource for practitioners, organizations, and regulators looking to understand and apply methods for robustness assessment in neural networks.
Key Topics
-
Definition of Robustness: The document defines robustness as the ability of an AI system to maintain its level of performance under any circumstances, especially when faced with unexpected or atypical data.
-
Assessment Workflow: The standard introduces a structured workflow for assessing robustness:
- Stating robustness goals
- Planning and executing tests
- Analyzing and interpreting outcomes
- Iterating based on results
-
Classification of Assessment Methods:
- Statistical Methods: Use mathematical testing on datasets to quantify performance under varying conditions.
- Formal Methods: Employ formal proofs and correctness checks to verify system properties within defined domains.
- Empirical Methods: Rely on experimentation, field trials, observation, and expert judgment to capture system behavior in real-world scenarios.
-
Robustness Metrics and Data Handling:
- Selection and sourcing of appropriate datasets, including considerations around diversity, representativeness, and outlier handling.
- Use of performance measures relevant to the application, such as root mean square error, accuracy, recall, and precision.
-
Recognized Limits and Challenges: The report acknowledges that complete characterization of robustness remains an open area of research, with both validation and testing methods facing practical limitations, especially due to the complexity and unpredictability of neural networks.
Applications
SIST-TP EN ISO/IEC/TR 24029-1:2024 is relevant to a wide range of stakeholders working with AI systems based on neural networks, including:
- Industry Practitioners: For evaluating the behavior and reliability of neural network-based applications in manufacturing, finance, and logistics.
- Regulatory Bodies: As a reference framework for robustness assessment in AI certification and quality assurance.
- Researchers: To gain an overview of state-of-the-art methodologies and identify areas for further investigation.
- Product Managers and Developers: For integrating robustness assessment into validation and verification processes of AI-powered products and services.
- Organizations in Safety-Critical Sectors: Such as automotive, healthcare, and aviation, where AI robustness is crucial for safety, compliance, and risk mitigation.
Related Standards
To provide a robust framework for AI system evaluation, the following related standards should be considered alongside SIST-TP EN ISO/IEC/TR 24029-1:2024:
- ISO/IEC 2382:2015 – Information Technology Vocabulary
- ISO/IEC 25000:2014 – Systems and Software Engineering - Systems and Software Quality Requirements and Evaluation (SQuaRE)
- ISO/IEC/IEEE 15288:2015 – Systems and Software Engineering - System Life Cycle Processes
- ISO/IEC/IEEE 26513:2017 – Systems and Software Engineering - Requirements for testers and reviews
Keywords: AI robustness assessment, neural network validation, artificial intelligence standards, neural network testing, robustness metrics, CEN artificial intelligence, formal methods, empirical methods, statistical methods, AI system reliability, neural network robustness, AI quality assurance, ISO/IEC TR 24029-1.