Research map / Federated learning and privacy / Privacy attacks and secure learning
INDEXGUARD: Index-only Backdoor Vetting for Secure Federated PEFT of Large Language Models
Javad Dogani, Devriş İşler, Nikolaos Laoutaris
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 |
|---|---|---|---|
| make label skew stronger (Dirichlet alpha 0.5 to 0.1) (ablation) | IndexGuard · AG News RIPPLES | detection accuracy: 99.1% -> 97.2% | Hurt |
| make label skew stronger (Dirichlet alpha 0.5 to 0.1) (ablation) | IndexGuard · AG News Syntactic | detection accuracy: 99.6% -> 98.7% | Hurt |
| make label skew stronger (Dirichlet alpha 0.3 to 0.1) (ablation) | IndexGuard · AG News InsertSent | detection accuracy: 100.00% -> 99.67% | No clear effect |
| use encoder-only BERT-base instead of Llama-2-7B (ablation) | IndexGuard · AG News InsertSent | AUC: 1.00 -> 0.996 | Hurt |
| use RoBERTa-base instead of Llama-2-7B (ablation) | IndexGuard · SST-2 InsertSent | AUC: 0.996 -> 0.991 | Hurt |
| swap LoRA for other PEFT methods (AdaLoRA, QLoRA, LoRA+, DoRA) (ablation) | IndexGuard · InsertSent AG News/SST-2 | AUC: all >= 0.998 | No clear effect |
| apply IndexGuard filtering of malicious clients (ablation) | federated PEFT · BERT-base SST-2 / Llama-2-7B AG News | ASR | Improved |
| apply IndexGuard filtering of malicious clients (ablation) | federated PEFT · SST-2 | clean accuracy: 0.921 -> 0.895 | Hurt |
| apply IndexGuard filtering of malicious clients (ablation) | federated PEFT · AG News | clean accuracy: 0.690 -> 0.651 | Hurt |
| attacker mixes more poisoned indices into reported support (b 0.1 to 0.5) (ablation) | IndexGuard · Adaptive reported-support attack | detection accuracy: 71.2 -> 100.0 | Improved |
| attacker mixes more poisoned indices (b 0.1 to 0.5) (ablation) | unmitigated attack · Adaptive reported-support attack | ASR@10: 55.1% -> 82.4% | Improved |
| IndexGuard (unsupervised, index-only) vs PEFTGuard | IndexGuard · SC average | AUC: 1.000 -> 0.998 | No clear effect |
Closest papers by meaning
- Forget to Flourish: Leveraging Machine-Unlearning on Pretrained Language Models for Privacy Leakage (AAAI 2025 · similarity 0.86)
- Certifying Language Model Robustness with Fuzzed Randomized Smoothing: An Efficient Defense Against Backdoor Attacks (ICLR 2025 · similarity 0.85)
- Privacy Backdoors: Enhancing Membership Inference through Poisoning Pre-trained Models (NeurIPS 2024 · similarity 0.85)
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.85)
- Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models (ICLR 2025 · similarity 0.85)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (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.
