Research map / Image generation / Text-to-image diffusion
Efficient Input-level Backdoor Defense on Text-to-Image Synthesis via Neuron Activation Variation
Shengfang Zhai, Jiajun Li, Yue Liu, Huanran Chen, Zhihua Tian, Wenjie Qu, Qingni Shen, Ruoxi Jia, Yinpeng Dong, Jiaheng Zhang
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 |
|---|---|---|---|
| use neuron activation variation instead of raw neuron coverage (ablation) | Ours · MS-COCO detection | AUROC: 54.5 -> 96.3 | Improved |
| use neuron activation variation instead of NC variation (ablation) | Ours · MS-COCO detection | AUROC: 71.1 -> 96.3 | Improved |
| use all layers instead of only down blocks (ablation) | Ours · Layer selection | AUROC: 88.0 -> 94.1 | Improved |
| use all layers instead of only mid blocks (ablation) | Ours · Layer selection | AUROC: 90.9 -> 94.1 | Improved |
| use all layers instead of only attention layers (ablation) | Ours · Layer selection | AUROC: 94.0 -> 94.1 | No clear effect |
| use all layers instead of only conv layers (ablation) | Ours · Layer selection | AUROC: 93.5 -> 94.1 | No clear effect |
| use first-step activations instead of full 50-step generation (vs UFID) | Ours · Detection time | seconds per sample: 33.041 -> 1.242 | Improved |
| compare with T2IShield on time cost | Ours · Detection time | seconds per sample: 7.445 -> 1.242 | Improved |
| compare against BadAct baseline | Ours · T2I backdoor detection | AUC avg: 60.7 -> 95.3 | Improved |
| apply to Diffusion Transformer (SD v3.5) vs UFID | Ours · RickBKD-TPA | AUROC: 48.3 -> 99.9 | Improved |
| reduce regularization weight alpha from 1e-5 to 1e-9 in adaptive attack (ablation) | Adaptive attack · MS-COCO | FID: 64.7 -> 13.2 | Mixed |
Closest papers by meaning
- Personalization as a Shortcut for Few-Shot Backdoor Attack against Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.89)
- TERD: A Unified Framework for Safeguarding Diffusion Models Against Backdoors (ICML 2024 · similarity 0.87)
- BackdoorDM: A Comprehensive Benchmark for Backdoor Learning on Diffusion Model (NeurIPS 2025 · similarity 0.87)
- Distraction is All You Need: Memory-Efficient Image Immunization against Diffusion-Based Image Editing (CVPR 2024 · similarity 0.87)
- Token by Token, Compromised: Backdoor Vulnerabilities in Unified Autoregressive Models (NeurIPS 2026 · similarity 0.87)
- Backdoor Attacks in Token Selection of Attention Mechanism (ICML 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.
