HRI Privacy

Wednesday, October 21, 2020

Problem statement

  • We want to balance usability and privacy for multi-user setting with a robot (conversational privacy).
  • Using Misty 2, which has a camera, 3 microphones, eye display, 2 speakers, 6 touch panels, neck/arm/body movement, wifi, bluetooth, bump sensors, Qualcomm snapdragon, windows, android

Contextual tuples

  • How do we connect and map these contexts? (e.g. deciding when to speak)
  • How do we map speakers to entities in the content? (e.g. Bob says: Alice should do this)
  • How do we determine how close users/speakers are to one another?
  • How do we do intent recognition? (question, answer, response, or exposition)
  • How do we determine where a conversation begins and ends?

Solutions

  • Recency bias
  • Context tuple matching system
  • Generative knowledge graph

Personalized permissions controller

  • Permissions controller which is enforced by privacy scores based on context tuples Context = (entities, sentiments, paralinguistic features, category)
  • Controller is personalized via adjusting thresholds/scores based on privacy indication phrases (“Feel free to tell X”, “Don’t tell anybody”)
  • Can we start with a generalized controller and fine-tune this to each user?
  • Should paralinguistic features also influence personalization?

Audio

  • How to map speakers and timestamps to text?
  • How to maintain paralinguistic features as well as content based features? How to add these to our context tuple?

Information Graph

  • How do we store conversations in a way that is easy to query and space conscious?
  • How do we map and link information that is related to each other?

Datasets

  • We need a privacy indication phrase dataset.
  • We need a privacy score initialization dataset.
  • Should we also make an inference dataset about conversations based on prior information or conversations?

User studies

  • What task should the robot do in the user study?

Language

  • Contextually similar people understand each other better. Updating each other with experiences and interesting conversations.
  • If you know someones encoding and decoding scheme for context and communication better, we can translate experiences easier.
  • Friend drifting apart is natural, unless we spend time and effort to update each other on significant experiences, feelings, and thoughts.
researchroboticsprivacynlp

Meditation, taiji, and mindset

HAI 2020 AI