Securing the ML Lifecycle
Figure 5-1 : Initial Security Assessment
Based on our discussions of ( selected ) attacks and countermeasures , we now present a checklist with sample questions that could be used ( and should be added to in any given organization ) to support initial discussions about securing a machine learning project .
Questions labelled “ standard ” should be asked in any software project , while “ advanced ” questions require in-depth technical knowledge of the ML pipeline .
Overall Project
Do you know your data assets ?
Question / Action Item
Do you know all involved stakeholders and where your ML pipeline is executed ? Can you clearly articulate your business objective ( needed to validate your model )? What interest could an attacker have in downgrading or evading your model ?
Do your project and the involved data have to meet any regulatory demands or compliance criteria ?
Table 5-1 : Overall Project Assessment
Standard Advanced
✓
✓
✓
✓
✓
48 March 2020