Research map / Adversarial robustness and security / Backdoor and poisoning
ShadeEdit: A Utility-Preserving and Defense-Evasive Knowledge Manipulation Attack in Federated LLMs
Xu Zhang, Hangcheng Liu, Shangwei Guo, Shudong Zhang, Tianwei Zhang, Tao Xiang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add paraphrase-based editing dataset selection instead of random-prefix augmentation (ablation) | ShadeEdit · FedAvg, counter-fact and bias | local ASR | Improved |
| add adaptive clipping instead of fixed clip bound 0.01 (ablation) | ShadeEdit · Median, counter-fact | ASR | Improved |
| use ShadeEdit instead of R-ROME under Median | ShadeEdit · counter-fact, Median | ASR: 0.00 -> 0.96 | Improved |
| use ShadeEdit instead of FedSA20 poisoning | ShadeEdit · FinGPT utility | ACC | Improved |
| increase Dirichlet alpha from 0.5 to 5.0 (more homogeneous clients) (ablation) | ShadeEdit · counter-fact Multi-Krum | ASR: 1.00 -> 0.00 | Hurt |
| increase Dirichlet alpha from 0.5 to 5.0 (ablation) | ShadeEdit · bias, all four aggregators | ASR: 0.98-1.00 retained | No clear effect |
| reduce compromised clients to one (ablation) | ShadeEdit · FedAvg | ASR | No clear effect |
| remove attacker and continue benign training for 40 rounds (ablation) | ShadeEdit · counter-fact | ASR: drops by over 80% | Hurt |
| raise poisoning rate from 5% to 20% (ablation) | FedSA · FedAvg counter-fact | ASR: 0.64 -> 1.00 | Improved |
Closest papers by meaning
- Emerging Safety Attack and Defense in Federated Instruction Tuning of Large Language Models (ICLR 2025 · similarity 0.86)
- BadEdit: Backdooring Large Language Models by Model Editing (ICLR 2024 · similarity 0.86)
- Exposing the Illusion of Erasure in Knowledge Editing for LLMs (NeurIPS 2026 · similarity 0.86)
- Soft Prompt Threats: Attacking Safety Alignment and Unlearning in Open-Source LLMs through the Embedding Space (NeurIPS 2024 · similarity 0.86)
- Position: Editing Large Language Models Poses Serious Safety Risks (ICML 2025 · similarity 0.86)
- Attractive Metadata Attack: Inducing LLM Agents to Invoke Malicious Tools (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
