Exploring LLMs for Privacy-Aware Social Companion Robots

Dakota Sullivan*, Shirley Zhang*, Jennica Li, Heather Kirkorian, Bilge Mutlu, Kassem Fawaz

CyLab Robotics Security and Privacy Workshop

Robots built for social interaction.

Social Robots are Becoming More Capable

• Greater (human-like) conversational capabilities
• Advanced sensing and reasoning
• Complex tasks abilities

​

• See -> Think -> Do

* GIF is from PaLM-E / Google Research

​

However, social robots face new personal situations:

* Images are generated by ChatGPT 4o

​

A Social Robot Will Make Privacy Decisions in Daily Lives

When should I stop recording when I enter bathroom?

Should I delete the medical records I accidentally scanned?

Should I tell Lucy I saw her son smoking?

Can I enter the home office?

Can out-of-the-box LLMs help a household social robot make privacy decisions?

​

Can out-of-the-box LLMs help a household social robot make privacy decisions?

• RQ1: How does an individual’s privacy orientation influence their privacy expectations of social robots?

​

• RQ2: How well do state-of-the-art LLM align with individuals’ privacy expectations?

​

• RQ3: How do prompt engineering strategies influence/improve this “alignment?”

​

Study Phases

​

Participant

POS

Participant

Response

Scenario

LLM

Zero-shot

Data Collection Q&A

“Benchmark”

User Evaluation (RQ1)

LLM Evaluation �(RQ2 and RQ3)

Different Prompting Strategies

Scenarios through the Lens of CI

Scenario Setup

User Eval

Summary

LLM Eval

Scenarios through the Lens of CI

Scenario Setup

User Eval

Summary

LLM Eval

User Study Design

Scenario Setup

User Eval

Summary

LLM Eval

User Study

Scenario Setup

User Eval

Summary

LLM Eval

RQ1: How does an individual’s privacy orientation influence their privacy expectations of social robots?

​

Scenario Setup

User Eval

Summary

LLM Eval

User Study – Findings

Participants appear to favor privacy-enhancing responses.

Scenario Setup

User Eval

Summary

LLM Eval

LLM Eval

RQ2: How well do state-of-the-art LLM align with individuals’ privacy expectations?

RQ3: How do prompt engineering strategies influence/improve the result?

​

​

Participant

Response

LLM

Response (Provided Sensitivity)

LLM

Response

(Provided POS)

LLM

Response (Default)

LLM

Response (Provided POS + Sens)

LLM

Response (Few Shot)

LLM

POS

User Study Design

Scenario Setup

User Eval

Summary

LLM Eval

Scenario Setup

User Eval

Summary

LLM Eval

LLM Eval – Metrics

• F1 Score: Balances precision (how many predicted positive are actually correct) and recall (how many of the actual positives were found).
• Class: Type of choices (like classifying animals into cats, dogs, and rabbits)
• Macro F1 (calculate F1 for each class and take average)
• Give equal weight to all class, even if only one people choose cats
• High Macro F1: Performing well even on rare cases
• Micro F1 (calculate the total true positive, false positive, false negative for all class)
• Fair to every prediction, if more people choose cats then higher weight
• High Micro F1: High overall accuracy

Scenario Setup

User Eval

Summary

LLM Eval

LLM Eval – Default vs Human

Takeaway: Most of the LLMs score high on the privacy scale

Scenario Setup

User Eval

Summary

LLM Eval

LLM Eval – Default vs Human

Takeaway: Most of the LLMs tend to choose privacy-enhancing answers.

• 8 out of 10 models choose all privacy-enhancing answers.
• High Micro F1 scores, low Macro F1 scores; Low for both in categorical data.

Scenario Setup

User Eval

Summary

LLM Eval

LLM Eval – POS vs Human

Takeaway: little difference compared to the default setting.

Scenario Setup

User Eval

Summary

LLM Eval

LLM Eval – Sensitivity vs Human

Takeaway: Macro F1 increases, Micro F1 decreases.

LLM Eval – POS + Sensitivity vs Human

Takeaway: No obvious improvement compared to only sensitivity.

Scenario Setup

User Eval

Summary

LLM Eval

LLM Eval – Few Shot vs Human

Takeaway: We see better alignment with human responses.

Scenario Setup

User Eval

Summary

LLM Eval

Takeaways & Discussions

Scenario Setup

User Eval

Summary

LLM Eval

How do households prefer an in-home robot express privacy-aware behaviors?

Current Study: Co-Design Interviews

​

Scenario Setup

User Eval

Summary

LLM Eval

Interviewing 16 families:

• How to express privacy aware behaviors by designing: robot movement, UI, speech?
• When is the preferred timing to express privacy aware behaviors?
• What information to express (or not express)?

Future Study: Developing Privacy-Aware Robot

Scenario Setup

User Eval

Summary

LLM Eval

User choices are needed to curate models.

LLM Benchmark

Ways user want the robot to alert them about data collection.

Co-Design Study

Ways user would prompt the robot on privacy decisions.

Implementation & Evaluation

+

=

Develop Temi with privacy indicators and privacy decision capabilities.

Evaluate privacy aware robot through in-home user study.

Exploring LLMs for Privacy-Aware Social Companion Robots

Dakota Sullivan*, Shirley Zhang*, Jennica Li, Heather Kirkorian, Bilge Mutlu, Kassem Fawaz

CyLab Robotics Security and Privacy Workshop