Data on Demand: The Challenges of Building a Privacy Focused AI Device
Details architectural and data pipeline practices used at Rabbit to secure customer identities, anonymize voice data, and control RAG model access.
Summary
Privacy Dilemmas in Consumer AI Devices
Building standalone AI hardware products requires collecting extensive user identity, audio, and personal context data, necessitating strong privacy-by-design architectures.
Identity Management and Credential Isolation
Designing authentication pipelines to collect and manage customer web credentials securely prevents lateral unauthorized data access.
Voice Telemetry Anonymization Pipelines
Processing, logging, and anonymizing customer voice interactions in real time prevents sensitive voice biometric leaks during model training.
Just In Time Context Access for RAG Models
Implementing strict 'just-in-time' access policies ensures retrieval-augmented generation (RAG) models only ingest contextual data when explicitly authorized by users.
Related talks
Creator Stage·26 min
Patrick Walsh · Bob Wall
Explains embedding inversion attacks against generative AI vector databases and introduces open-source vector encryption for private search.
Creator Stage·26 min
Chelsea Button · Alfonso Tinoco · Elaine Shi
Explains oblivious computation algorithms to hide access patterns in encrypted databases, with applications across blockchains, contact discovery, and AI.
Creator Stage·26 min
Rachel Cummings
Examines operational challenges in deploying differential privacy systems across real-world enterprise environments and government agencies.