Research map / Adversarial robustness and security / Backdoor and poisoning
Practical Poisoning Attacks on Neural Networks
Junfeng Guo, Cong Liu
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 a pre-trained model P much more accurate than the target model T (ablation) | BlackCard · CIFAR-10 (multi-shot) | misclassification confidence: 100% -> 0.15% | Hurt |
| use a pre-trained model P much more accurate than the target model T (ablation) | BlackCard · MNIST (multi-shot) | misclassification confidence: 100% -> 6.43% | Hurt |
| use a pre-trained model P much more accurate than the target model T (ablation) | BlackCard · Fashion-MNIST (multi-shot) | misclassification confidence: 100% -> 11.73% | Hurt |
| use P with similar or lower accuracy than T (ablation) | BlackCard · MNIST, Fashion-MNIST, CIFAR-10 | misclassification confidence: 8.21% -> 100% | Improved |
| use a P about 10% less accurate than T (ablation) | BlackCard · CIFAR-10 (multi-shot) | misclassification confidence: 100% | No clear effect |
| increase the number of poison samples (ablation) | BlackCard · six tasks | misclassification confidence: 7.53% -> 100% | Improved |
| switch from one-shot to multi-shot poisoning | BlackCard · CIFAR-10 | attack success rate: 10.4% -> 100% | Improved |
| add the feature-space collision-avoidance term for the base instance in the poison objective (ablation) | BlackCard · analysis | misclassification confidence | Improved |
| use a pre-trained model with different architecture than the target (DenseNet121 vs ResNet) | BlackCard · CIFAR-10 | attack success rate: 100% | No clear effect |
Closest papers by meaning
- Witches' Brew: Industrial Scale Data Poisoning via Gradient Matching (ICLR 2021 · similarity 0.85)
- Robust Learning for Data Poisoning Attacks (ICML 2021 · similarity 0.84)
- A Theoretical Analysis of Backdoor Poisoning Attacks in Convolutional Neural Networks (ICML 2024 · similarity 0.84)
- Sharpness-Aware Data Poisoning Attack (ICLR 2024 · similarity 0.83)
- Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples (NeurIPS 2026 · similarity 0.81)
- Attack To Defend: Exploiting Adversarial Attacks for Detecting Poisoned Models (CVPR 2024 · similarity 0.81)
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.
