Overview
ISO/IEC 24029-2:2023 - "Artificial intelligence (AI) - Assessment of the robustness of neural networks - Part 2: Methodology for the use of formal methods" provides a structured methodology for applying formal methods to prove robustness properties of neural networks. The standard guides how to select, apply and manage formal verification techniques (e.g., solvers, abstract interpretation, reachability analysis, model checking) to obtain mathematical proofs, counterexamples or inconclusive results about network behavior. It is intended to complement statistical testing and increase trust in neural network robustness across the AI life cycle.
Key topics and technical requirements
- Robustness assessment framework: Definitions and criteria for robustness-related properties including stability, sensitivity, relevance and reachability, and how to express these as formal requirements.
- Domain and attributes: Guidance on defining the input domain for verification (attributes, bounding conditions) so formal analysis targets relevant operating regions.
- Applicability of formal methods: Considerations for different neural network architectures and input data types, and mapping of techniques to tasks:
- Solvers (SMT/LP/MILP) for exact reasoning where feasible
- Abstract interpretation for scalable over-approximations
- Reachability analysis (deterministic and non-deterministic environments)
- Model checking for stateful or symbolic models
- Life-cycle integration: Recommendations for applying formal methods during design & development, verification & validation, deployment and operation/monitoring - including identifying recognized features, checking separability, covering input domains, and measuring perturbation impact.
- Practical constraints: Notes on scalability and trade-offs between precision and computational cost; formal methods may be complementary rather than universally decisive.
Practical applications and intended users
ISO/IEC 24029-2:2023 is targeted at:
- AI developers and ML engineers seeking formal assurance of neural network behavior.
- Verification and safety engineers in safety-critical domains (autonomous driving, medical devices, industrial control).
- Regulatory and compliance teams needing documented, rigorous robustness evidence.
- Researchers and tool developers building formal verification tools for neural networks.
Typical applications:
- Proving that small input perturbations do not change classifier labels (adversarial robustness).
- Verifying reachability properties for control systems.
- Complementing statistical testing during V&V and monitoring robustness changes in operation.
Related standards
Keywords: ISO/IEC 24029-2:2023, formal methods, robustness assessment, neural networks, formal verification, model checking, abstract interpretation, reachability analysis.
