Is your application designed in a secure way? Can you trace down who performed which actions on your pipeline, model versions, or data connections?
Why audit your models?
Algorithms may influence if we get hired for a job, what college we attend and even our political views. With AI making more and more decisions that affect our daily lives we need to look critically at how these applications are designed and operated. This is where AI auditing will play an important role. One of the goals of AI Auditing is to gain insight in and track which person did what adjustment to a specific model (version). This way you can prevent any kind of misuse and stay in full control of your application. Most importantly, you create more trust for the end-user and other responsible parties.
How does UbiOps help you with this?
UbiOps introduced the ‘audit event’ feature where you can track who performed what change to which model (version), when and where. Furthermore, you can ensure that no one uploads a different version of your ML model or connects the pipeline to a different database.
Using this feature it is very easy to investigate and verify that all operations were executed as planned. You can do so via the UI, the CLI or the client library.
Further research on AI auditing is needed
AI auditing is still very much in its early stages and audit events are just one step forward in the auditing process. Because it’s an emerging topic, no established methods have been tested yet to check if the AI model meets certain requirements. However, taking this first step and continuing the discussion with our network partners such as PwC is crucial to help in building a trustworthy framework for AI audits.
Is your company already actively auditing models? Then we are curious to hear from you how that is going and how you have tackled it.
“Working towards a transparent and controlled approach to the design, implementation and operation of AI application is critical, as AI will play an increasing role in our society. UbiOps takes the first steps in enabling the much needed trust”.
Mona de Boer, partner Data & Technology, PwC Netherlands
