These principles help developers, engineers, decision makers and risk owners make informed decisions about the design, development, deployment and operation of their machine learning (ML) systems.
Pages
Page 9 of 22
2.2 Secure your development infrastructure
Goals:
You understand that datasets, models and other artefacts are crucial assets that need to be protected.
You understand the potential impact of any theft or loss of these assets (including legal and reputational impacts).
You know what data you hold, where it's stored and how it is being used.
Why this is this important?
Like your data and models, your development infrastructure needs to be sourced through a trusted supply chain, whether you choose to use local compute or public cloud services. Although infrastructure security is not unique to ML, it is particularly important in ML due to the possibility that a compromise at this stage could impact throughout a system's life cycle. Your digital assets represent a significant investment of resources and intellectual property, making them an attractive target for theft, while their manipulation could significantly affect your system's operation.
Securing your development infrastructure is likely to involve assessing multiple components and settings, applying fundamental cyber security best practice. It's important to ensure that software and operating systems are updated to the latest versions and that access to your environment is limited to those with a legitimate need. Access to and modification of components should be logged and monitored. Your specific use case will influence your security requirements, but you should understand that initiating development in a less secure R&D environment could result in vulnerabilities that are hard to remove if you want to move a product into production.
How could this principle be implemented?
Follow cyber security best practice and recognised IT security standards
The importance of following cyber security best practice is not unique to ML development. Sensible steps to take include protecting data at rest and protecting data in transit, patching software to the latest versions, implementing multi-factor authentication, security logging and, where necessary, using two-person controls. ISO/IEC 27001 is a globally used standard for information security management systems that promotes a holistic approach to managing cyber risks.
Access to your training environment should follow the principle of least privilege for user access. Implement access logging and monitoring in your development environment. The NCSC has guidance about logging for security purposes, and CISA now maintains Logging Made Easy software (originally developed by the NCSC).
Use secure software development practices
Secure software development practices provide steps to safeguard your software development life cycle and ways to improve the security of your ML model and/or system. See the NCSC's guidance on secure development and deployment and CISA’s Secure by Design for further guidance.
Monitor common vulnerabilities associated with your development software and any code libraries you rely on. The common vulnerabilities and exposures (CVE) program provides a glossary (maintained by the MITRE corporation) that identifies, defines and catalogues publicly disclosed cyber security vulnerabilities.
Be aware of legal and regulatory requirements
Ensure decision makers understand what data you hold, and the law and regulations around data collection. Ensure your developers understand the impact and consequences of data breaches, their responsibility when processing data and the importance of developing secure software. The NCSC has guidance on approaches to securing personal data securely and the Information Commissioner's Office (ICO) publish information on the laws surrounding data protection.
