Get Started
Home
Topics
Search
Library
Research questionHow can split-LLM training prevent returned gradients from revealing which rows are real?Forward-channel privacy checks can pass even when gradient patterns reveal which received rows carried the real computation. This hidden signal can expose frame structure and provide information about private contents.
AI
Alignment & Safety
LLM Pretraining & Post-training
Machine Learning
Latest papersRecent research connected to this question, newest first.Privacy Failure in Split-LLM Training, The Returned Gradient Nullifies the DecoysThe evidence concerns a two-node system where a trusted local node holds the private loss and returns output gradients to an untrusted cloud node. Real and decoy rows are mixed in each frame, and the reported finding covers row-wise gradient clipping and noising, while attacks accumulated across training steps and four other attack classes were not measured.research paper · Sep 3, 2026
Related questions
How can LLM pretraining avoid sudden gradient explosions when scaling to larger models?How can black-box LLMs resist jailbreaks without weight access or retraining while preserving benign-query utility?How can LLMs remove undesirable knowledge while preserving utility with limited retention and unlearning data?How can differentially private federated learning preserve faithful explanations without sacrificing predictive utility?