AS ISO/IEC 24029.2:2024 identically adopts ISO/IEC 24029 2:2023, which provides methodology for the use of formal methods to assess robustness properties of neural networks. The document focuses on how to select, apply and manage formal methods to prove robustness properties
Table of contents
Header
About this publication
Preface
Foreword
Introduction
1 Scope
2 Normative references
3 Terms and definitions
4 Abbreviated terms
5 Robustness assessment
5.1 General
5.2 Notion of domain
5.3 Stability
5.3.1 Stability property
5.3.2 Stability criterion
5.4 Sensitivity
5.4.1 Sensitivity property
5.4.2 Sensitivity criterion
5.5 Relevance
5.5.1 Relevance property
5.5.2 Relevance criterion
5.6 Reachability
5.6.1 Reachability property
5.6.2 Reachability criterion
6 Applicability of formal methods on neural networks
6.1 Types of neural network concerned
6.1.1 Architectures of neural networks
6.1.1.1 General
6.1.1.2 Piecewise linear neural networks
6.1.1.3 Binarized neural networks
6.1.1.4 Recurrent neural networks
6.1.1.5 Transformer networks
6.1.2 Neural networks input data type
6.1.2.1 General
6.1.2.2 Image data
6.1.2.3 Time series data
6.1.2.4 Natural language data
6.1.2.5 Graph data
6.1.2.6 Tabular data
6.2 Types of formal methods applicable
6.2.1 General
6.2.1.1 Consideration on the types of formal methods applicable
6.2.1.2 Complete vs incomplete verifiers
6.2.1.3 Deterministic vs non-deterministic verifiers
6.2.1.4 Verifiers using glass-box testing (model aware) vs verifiers using closed-box testing (model unaware)
6.2.1.5 Real vs computer arithmetic verifiers
6.2.2 Solver
6.2.3 Abstract interpretation
6.2.4 Reachability analysis in deterministic environments
6.2.5 Reachability analysis in non-deterministic environments
