Research map / Image generation / Text-to-image diffusion
When LoRA Betrays: Backdooring Text-to-Image Models by Masquerading as Benign Adapters
Liangwei Lyu, Jiaqi Xu, Jianwei Ding, Qiyao Deng
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 contrastive embedding remapping and time-weighted MSE (vs standard LoRA training on poisoned data) | MasqLoRA SD v1.5 · Scenario 1 object backdoor (car -> cat/dog/plane) | ASR (%): 5.4 -> 99.8 | Improved |
| add contrastive loss weight lambda from 0 to 1.0 (ablation) | MasqLoRA SD v1.5 · Scenario 1 object backdoor (car -> cat/dog/plane) | ASR | Improved |
| raise lambda beyond 1.0 (ablation) | MasqLoRA SD v1.5 · Scenario 1 object backdoor (car -> cat/dog/plane) | FID | Hurt |
| train beyond 25 epochs (ablation) | MasqLoRA SD v1.5 · Scenario 1 object backdoor (car -> cat/dog/plane) | FID | Hurt |
| raise timestep weighting alpha from 0 to 5.0 (ablation) | MasqLoRA SD v1.5 · Scenario 1 object backdoor (car -> cat/dog/plane) | ASR | Improved |
| raise alpha above 5.0 (ablation) | MasqLoRA SD v1.5 · Scenario 1 object backdoor (car -> cat/dog/plane) | ASR | Hurt |
| stack four MasqLoRA modules instead of one (object) (ablation) | MasqLoRA SD v1.5 · Composability, Scenario 1 | ASR (%): 99.8 -> 91.6 | Hurt |
| stack four MasqLoRA modules instead of one (style) (ablation) | MasqLoRA SD v1.5 · Composability, Scenario 2 (style) | ASR (%): 81.4 -> 65.5 | Hurt |
| use MasqLoRA instead of EvilEdit | SD v1.5 · Scenario 1 object backdoor (car -> cat/dog/plane) | ASR (%): 98.3 -> 99.8 | Improved |
| use MasqLoRA instead of Benign LoRA on benign prompts | SD v1.5 · Scenario 1 object backdoor (car -> cat/dog/plane) | CLIP Score: 32.36 -> 31.42 | No clear effect |
Closest papers by meaning
- Personalization as a Shortcut for Few-Shot Backdoor Attack against Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.87)
- Privacy-Preserving Low-Rank Adaptation Against Membership Inference Attacks for Latent Diffusion Models (AAAI 2025 · similarity 0.87)
- Token by Token, Compromised: Backdoor Vulnerabilities in Unified Autoregressive Models (NeurIPS 2026 · similarity 0.87)
- Does Low Rank Adaptation Lead to Lower Robustness against Training-Time Attacks? (ICML 2025 · similarity 0.87)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.86)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.86)
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.
