TriShield: Zero-Utility-Loss Defense Against Privacy Backdoors in Federated Language Model Fine-Tuning via Orthogonal Gradient Projection and Optimizer State Entanglement
Cheng Wei
Honor Device Co., Ltd., Shenzhen, China
cs.LG, cs.CL
Submitted: 2026-07-30
Comments: 12 pages,3 figures
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Terminology
Sources
- From Efficiency to Leakage -- Privacy Backdoor in Federated Language Model Fine-Tuning
- LoRA: Low-Rank Adaptation of Large Language Models
- Gradient Projection Memory for Continual Learning
- FLAME: Taming Backdoors in Federated Learning (Extended Version 1)
- Inverting Gradients -- How easy is it to break privacy in federated learning?
- iDLG: Improved Deep Leakage from Gradients
- How To Backdoor Federated Learning
- InstaHide: Instance-hiding Schemes for Private Distributed Learning
- Ensemble Distillation for Robust Model Fusion in Federated Learning
- Federated Learning Based on Dynamic Regularization
- When the Curious Abandon Honesty: Federated Learning Is Not Private
- PRECODE - A Generic Model Extension to Prevent Deep Gradient Leakage
- Gradient Projection For Continual Parameter-Efficient Tuning
- R-GAP: Recursive Gradient Attack on Privacy
Related papers
- Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
- AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
- Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
- Convergence issues in Relational Concept Analysis based on AOC-posets
- Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks