# TriShield: Zero-Utility-Loss Defense Against Privacy Backdoors in Federated Language Model Fine-Tuning via Orthogonal Gradient Projection and Optimizer State Entanglement

arXiv: https://arxiv.org/abs/2607.27940 · HTML: https://mlsanity.com/paper/2607.27940/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Federated unlearning and poisoning, one of 12 directions in LLM safety and security.
- Builds on 16 of its 22 references that are in the map.
- First posted on arXiv 30 Jul 2026.

## Placed in

- LLM safety and security / [Federated unlearning and poisoning](https://mlsanity.com/safety/federated-unlearning-and-poisoning/)

## Builds on (most cited first, 16 of 16)

- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [Communication-Efficient Learning of Deep Networks from Decentralized Data](https://mlsanity.com/paper/1602.05629/) (2016)
- [Prefix-Tuning: Optimizing Continuous Prompts for Generation](https://mlsanity.com/paper/2101.00190/) (2021)
- [Deep Leakage from Gradients](https://mlsanity.com/paper/1906.08935/) (2019)
- [How To Backdoor Federated Learning](https://mlsanity.com/paper/1807.00459/) (2018)
- [Ensemble Distillation for Robust Model Fusion in Federated Learning](https://mlsanity.com/paper/2006.07242/) (2020)
- [Federated Learning Based on Dynamic Regularization](https://mlsanity.com/paper/2111.04263/) (2021)
- [iDLG: Improved Deep Leakage from Gradients](https://mlsanity.com/paper/2001.02610/) (2020)
- [Gradient Projection Memory for Continual Learning](https://mlsanity.com/paper/2103.09762/) (2021)
- [Evaluating Gradient Inversion Attacks and Defenses in Federated Learning](https://mlsanity.com/paper/2112.00059/) (2021)
- [When the Curious Abandon Honesty: Federated Learning Is Not Private](https://mlsanity.com/paper/2112.02918/) (2021)
- [R-GAP: Recursive Gradient Attack on Privacy](https://mlsanity.com/paper/2010.07733/) (2020)
- [InstaHide: Instance-hiding Schemes for Private Distributed Learning](https://mlsanity.com/paper/2010.02772/) (2020)
- [FL-WBC: Enhancing Robustness against Model Poisoning Attacks in Federated Learning from a Client Perspective](https://mlsanity.com/paper/2110.13864/) (2021)
- [PRECODE - A Generic Model Extension to Prevent Deep Gradient Leakage](https://mlsanity.com/paper/2108.04725/) (2021)
- [From Efficiency to Leakage -- Privacy Backdoor in Federated Language Model Fine-Tuning](https://mlsanity.com/paper/2606.20553/) (2026)

## Abstract (arXiv)

Federated fine-tuning of large language models (LLMs) enables collaborative training without exposing raw data. However, a recent attack, NeuroImprint, demonstrates that a malicious parameter server can corrupt a PEFT adapter into a privacy backdoor: by assigning a dedicated memorization neuron to each training sample and ensuring each neuron updates at most once, the server can analytically reconstruct 59%--79% of client training data with high semantic fidelity. Existing defenses---including local differential privacy (LDP) and gradient clipping---either fail against this attack or impose unacceptable utility degradation. We present TriShield, a three-layer deterministic defense that completely prevents NeuroImprint-style reconstruction with zero model utility loss and no additional communication rounds. TriShield consists of: (1) a Parameter Artifact Detector that identifies memory-neuron signatures in distributed model parameters before local training begins; (2) a Stateful Virtual Iteration mechanism that forces Adam/AdamW's momentum state to irreversibly entangle gradients across virtual steps, invalidating NeuroImprint's closed-form inversion; and (3) a Zero-Utility Orthogonal Projection operator that projects all local gradient updates onto the main-task semantic subspace computed via SVD, physically eliminating any gradient components that carry private memorization. We prove theoretically that after Layers 2 and 3, the mutual information between the uploaded gradient and any individual training sample is zero. Experiments on GPT-2 (117M) and Llama-Guard-3-1B verify that TriShield reduces NeuroImprint reconstruction rate to 0% across all tested attack variants, while maintaining or improving training accuracy, with less than 5% additional GPU computation overhead.