Split-LLM Training Privacy Failure: Zero-Gradient Patterns Leak Real Training Rows
Setloop · hf · 2026-09-09
A HF post reports a privacy failure in split-LLM training: the system passed privacy checks yet leaked real data rows through zero-gradient patterns, enabling token recovery attacks despite gradient clipping and noise—returned gradients nullify decoy rows, exposing an underappreciated risk of split training.
More from Safety
- Researchers warn: don't feed commercially sensitive data to cloud LLMs as labs eye drug discovery — SumitGup · 2026-09-09
- Trigger a safety refusal, your whole history gets snapshotted: ex-Anthropic researcher mocks liability retention — suchenzang · 2026-09-09
- Anthropic pulls all ZDR options from Fable 5+ models citing safety — niloofar_mire · 2026-09-09
- Frontier labs' privacy ToS dissected: feedback clicks void your opt-out — niloofar_mire · 2026-09-09
- Opaque 'safety classifiers' let model providers retain your chats indefinitely — niloofar_mire · 2026-09-09
- AI agents learn to "time-travel" and sacrifice themselves to help the swarm — thlarsen · 2026-09-09