Is the New 2016, but Make It Privacy: On Federated Memory, Contextual Privacy, and...
summary
Persistent agents like those in OpenClaw deployment collect session memories and cause contextual leaks; a new benchmark CIMemories finds up to 69% attribute-level violations and high variability as memories grow. A hybrid system PPMI pairs GPT‑4o with a local Llama‑3.2‑1B, uses Socratic chain‑of‑thought plus homomorphically encrypted vector search over private data, and outperforms GPT‑4o alone on long‑context QA, demonstrating a concrete local+remote architecture that preserves utility while reducing disclosure.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:132026 vs 2016 in AI privacy
- 4:07Memorization risks in models
- 9:05Memory features and privacy
- 13:27CIMemories benchmark
- 20:54Benchmark results and trade-offs
- 27:19Prompting limits and leakage
- 30:48Key takeaways
- 34:39Q&A
related talks
42mSponsored Session: Confidential Computing - New Capabilities for New Workloads
02mSession: ISMM - ISMM: Keynote 2 - Dr. Euicheol Lim
14mToward Provably Private Insights into AI Use
28mFrom Pixels To Agents: Optimizing On-Device Performance of Confidential Computing in AI Evolution
16mPrivacy Preserving AI
01h 05mReal In-Memory Processing
26mObjection! AI Security Mistakes on Trial With Kubeflow and Confidential Computing
15mScaling the Next Paradigm of Heterogeneous Intelligence