Week 7 Discussion Topic: Access Control with AI

Goal for today: Design a new system for Access Control for shared files, built on LLMs

The university has just created a new file sharing system, to allow students (and professors) to share files in groups. This is intended to support classes (share files with the class?), group work (share files with your small group), student groups, faculty research groups, etc.

There need to be a way to specify who is allowed to see what.

New generative AI tools seem like a good way to do this? Your goal is to try to think through how this might work.

Requests

Normally, we would begin by collecting user data. We would find a bunch of (potential) users and ask them for examples. However, we are doing this in class, so we don’t have the time or ability to ask real users. Fortunately, YOU are in the target audience. So we can use some of your experience.

Start by thinking back to group sharing situations you’ve been in – sharing files with friends for homework; sharing files across class-assigned groups for group projects; sharing files with professors; sharing files with family back home. Be specific; make a (short) list of specific situations where you had to do this.

Now, think about the access control needs. What did you want to do, access control wise? Who did you want to have access? Who did you NOT want to have access? Did this change over time, or was it the same the whole time?

Make a specific, detailed list of example access control needs. Don’t try to be more specific than you were at the time; just write down what was going through your head at the time. Yes, those lists will not cover everything. That’s OK.

Permissions

Look through your list. What kinds of permissions did you want?

  • Read (View)
  • Read/Write
  • Delete
  • Edit?
  • Move
  • Rename
  • Share with others?
  • Organize / create subfolders?
  • Share Location
  • Access whole folder
  • Access specific photos / files

Make a list of what kinds of permissions you want your new system to support. Do you want to have all of them as separate permissions, or do you want to group them together into a smaller, easier to understand list? Use your list of examples to help you think not only about what permissions are needed, but how they are phrased and grouped.

Users

What users are needed? How many?

What kinds of users? Can we identify “roles” or “groups”?

Make a list. What users/groups are represented in your example list? How did the user (you) think about them? Was it individual people? Roles? Names, pre-existing groups? Unnamed, ad-hoc groups?

(I recommend creating some “fake users” to help you think through the interfaces. Give people names, groups as examples of ones that might need to be created.) Do you want to explicitly have roles or groups (like in file systems), or do you want to just list users and have the groups be implicit (like in Box/Dropbox)?

Objects

What are we talking about? Files? Programs to run? Types of requests?

What can people do with them? What might they want to do? What might they want to restrict?

Describe what kinds of things people might want to share in this kind of system. Be specific, using your list of examples.

Story

Now, let’s create a NEW persona-like example story. The goal of the story is to create a fake situation that captures all of the types of real examples from your list, but all in one story with a fake but detailed set of people, files, and actions.

Imagine one or more groups of people. Who are they? Give them a reason to need to share files. What are they trying to do? You’ll also need some people who don’t have access. What are the kinds of files / objects that they are going to share? Tell a story – what do they do with them? How might the access control needs change over time?

Chatbot Interface

Design the interface. What kinds of requests are needed? What kind of feedback is needed?

Write up some example access control requests. Use your story to provide examples / context for what needs to be done. Write up some chatbot examples of how the users in the story might be able to use your chatbot interface to set the access control settings appropriately. Try to figure out what you HOPE the chatbot does – how should it translate those requests from the user into actual access control lists / rules?

What are some problems that might come up? What happens if the chatbot gets it wrong? What happens when people aren’t specific enough for the chatbot? What does the chatbot need to know?

Note: SPARCLE

This idea we are working on is not new. For example, 20 years ago, IBM was working on a similar natural language interface for access control, called SPARCLE. You can see designs for it here.