Research map / Federated learning and privacy / Privacy attacks and secure learning
ShadowFPT: Backdooring Federated Prompt Tuning via Shadow Triggers
Kun Zhai, Teng Li, Yunhao Feng, Xingjun Ma
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ShadowFPT, a backdoor attack on federated prompt tuning of vision-language models in which a malicious client exploits the shared frozen backbone to implant backdoor behavior through the representation space while keeping its updates ordinary-looking.
Problem
Federated prompt tuning freezes a shared pretrained vision-language backbone and trains only lightweight prompts across clients, which creates a backdoor risk that existing defenses and attacks may not address.
Approach
Since the backbone is shared and frozen, a malicious client induces backdoor behavior via the representation space rather than through conventional poisoned prompt updates, while keeping its updates looking benign. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Shadow Trigger pretraining and injection (ShadowFPT-II) versus strongest baseline | federated prompt tuning backdoor · Caltech101 FPT-L | ASR (%): 19.81 -> 90.36 | Improved |
| pretrain trigger with auxiliary public data (ShadowFPT-II) instead of local data (ShadowFPT-I) | ShadowFPT · 9 dataset-setting pairs | ASR (%): 95.63 -> 97.67 | Improved |
| use a pretrained Shadow Trigger instead of a random trigger (ablation) | ShadowFPT · Caltech101 FPT-L | ASR (%): 21.32 -> 49.75 | Improved |
| add backdoor injection in FPT with adapted trigger (ablation) | ShadowFPT · Caltech101 FPT-L | ASR (%): 49.75 -> 90.36 | Improved |
| add layer-wise visual alignment to target-text alignment (ablation) | Shadow Trigger pretraining · Caltech101 FPT-L | ASR (%): 66.38 -> 90.36 | Improved |
| add target-text alignment to layer-wise alignment (ablation) | Shadow Trigger pretraining · Caltech101 FPT-L | ASR (%): 59.25 -> 90.36 | Improved |
| apply ShadowFPT-II under robust aggregation/defenses (Krum, LaSA, FLAME, CRFL) | ShadowFPT · defense rules | ASR gain over best baseline: +73.97, +76.55, +84.75, +74.55 | Improved |
| ShadowFPT-II across prompt modalities FPT-L/V/VL | ShadowFPT · Avg over datasets | ASR (%): 27.65 -> 96.42 (FPT-L) | Improved |
| ShadowFPT-II on fine-grained Oxford Flowers with textual prompts | ShadowFPT · Oxford Flowers FPT-L | clean accuracy (%): 66.63 -> 60.85 | Hurt |
| ShadowFPT-II clean-accuracy cost across defenses | ShadowFPT · Avg over datasets and settings | clean accuracy (pp): -1.71 average | Mixed |
Closest papers by meaning
- SABRE-FL: Selective and Accurate Backdoor Rejection for Federated Prompt Learning (ICLR 2026 · similarity 0.91)
- Exposing Functional Fusion: A New Class of Strategic Backdoor in Dynamic Prompt Architectures (CVPR 2026 · similarity 0.89)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.88)
- Leveraging Soft Prompts for Privacy Attacks in Federated Prompt Tuning (NeurIPS 2026 · similarity 0.87)
- Defending Multimodal Backdoored Models by Repulsive Visual Prompt Tuning (NeurIPS 2025 · similarity 0.86)
- VLMGuard: Bootstrapping Malicious Prompt Detectors from Unlabeled Vision-Language Prompts in the Wild (NeurIPS 2026 · 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.
