Research map / Adversarial robustness and security / Backdoor and poisoning
UNICORN: A Unified Backdoor Trigger Inversion Framework
Zhenting Wang, Kai Mei, Juan Zhai, Shiqing Ma
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 input-space transformations to represent triggers in different spaces instead of restricting inversion to pixel-space patterns | UNICORN · CIFAR-10, Patch, Blend, SIG, 1977 Filter, Kelvin Filter, Moon Filter, WaNet, and BppAttack | ASR-Inv: UNICORN: 99.61%, 99.74%, 99.17%, 96.54%, 94.57%, 97.50%, 99.95%, 99.33%; strongest baseline: 94.57%, 91.74%, 91.32%, 69. | Improved |
| Add the disentanglement constraint to separate compromised activations from benign activations (ablation) | UNICORN · CIFAR-10, attacks across pixel, signal, feature, and numerical spaces | ASR-Inv: average ASR-Inv is 94.25% with and 64.56% without disentanglement | Improved |
| Add the stealthiness constraint using SSIM (ablation) | UNICORN · Inverted backdoor samples | SSIM | Improved |
| Add the disentanglement constraint to Neural Cleanse (ablation) | Neural Cleanse on CIFAR-10, 1977 Filter, ResNet18 · CIFAR-10, 1977 Filter | ASR-Inv: 63.63% -> 70.35% | Improved |
| Add the disentanglement constraint to Neural Cleanse (ablation) | Neural Cleanse on CIFAR-10, 1977 Filter, ResNet18 · CIFAR-10, 1977 Filter | detection accuracy: 52.5% -> 55.0% | Improved |
| Vary alpha, beta, gamma, delta, and the clean-sample proportion (ablation) | UNICORN · CIFAR-10, Patch, 1977 Filter, WaNet, and BppAttack | SIM | Mixed |
Closest papers by meaning
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.89)
- DISTIL: Data-Free Inversion of Suspicious Trojan Inputs via Latent Diffusion (ICCV 2025 · similarity 0.89)
- Versatile Backdoor Attack With Visible, Semantic, Sample-Specific and Compatible Triggers (TPAMI 2025 · similarity 0.88)
- BAN: Detecting Backdoors Activated by Adversarial Neuron Noise (NeurIPS 2024 · similarity 0.88)
- LOTUS: Evasive and Resilient Backdoor Attacks through Sub-Partitioning (CVPR 2024 · similarity 0.88)
- Rethinking Backdoor Attacks on Dataset Distillation: A Kernel Method Perspective (ICLR 2024 · 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.
