Research map / Adversarial robustness and security / Backdoor and poisoning
Poisoned Distillation: Injecting Backdoors into Distilled Datasets Without Raw Data Access
Ziyuan Yang, Ming Yan, Yi Zhang, Joey Tianyi Zhou
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 |
|---|---|---|---|
| raise number of reconstructed archetypes m from 5 to 20 (ablation) | Poisoned distillation attack · CIFAR-10 IPC1 (DC) | benign accuracy (%): 25.79 -> 25.36 | No clear effect |
| raise number of reconstructed archetypes m from 5 to 20 (ablation) | Poisoned distillation attack · CIFAR-10 IPC1 (DC) | attack time (s): 41.5 -> 125.60 | Hurt |
| increase balance weight alpha toward backdoor loss (ablation) | Poisoned distillation attack · CIFAR-10 IPC1 (DC) | benign accuracy / ASR | Mixed |
| train attacker-side model as AlexNet instead of ConvNet (ablation) | Poisoned distillation attack · CIFAR-10 IPC10 (DC) | attack success rate (%): 100.00 -> 23.04 | Hurt |
| train attacker-side model as VGG11 instead of ConvNet (ablation) | Poisoned distillation attack · CIFAR-10 IPC10 (DC) | attack success rate (%): 100.00 -> 68.89 | Hurt |
| train attacker-side model as VGG11 instead of ConvNet (ablation) | Poisoned distillation attack · CIFAR-10 IPC10 (MTT) | attack success rate (%): 100.00 -> 77.21 | Hurt |
| increase user training epochs from 10 to 100 | Poisoned distillation attack · CIFAR-100 IPC1 (DC) | attack success rate (%): 22.56 -> 99.83 | Improved |
| inject backdoor into benign distilled set (benign accuracy cost) | Poisoned distillation attack · CIFAR-10 IPC10 (DC) | benign accuracy (%): 42.63 -> 35.92 | Hurt |
| inject backdoor into benign distilled set (benign accuracy cost) | Poisoned distillation attack · MNIST IPC1 (DC) | benign accuracy (%): 88.81 -> 79.45 | Hurt |
| user trains ResNet18 instead of ConvNet | Poisoned distillation attack · CIFAR-100 IPC1 (DC) | attack success rate (%): 99.83 -> 0.03 | Hurt |
| user trains ResNet18 instead of ConvNet | Poisoned distillation attack · CIFAR-10 IPC10 (DC) | attack success rate (%): 100.00 -> 88.00 | Hurt |
| apply attack to MTT-distilled data with AlexNet user instead of ConvNet | Poisoned distillation attack · CIFAR-10 IPC10 (MTT) | attack success rate (%): 100.00 -> 54.31 | Hurt |
Closest papers by meaning
- Rethinking Backdoor Attacks on Dataset Distillation: A Kernel Method Perspective (ICLR 2024 · similarity 0.92)
- Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized Scaled Prediction Consistency (ICLR 2024 · similarity 0.90)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.89)
- From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks (AAAI 2024 · similarity 0.89)
- Taught Well Learned Ill: Towards Distillation-conditional Backdoor Attack (NeurIPS 2025 · similarity 0.88)
- Privacy Backdoors: Enhancing Membership Inference through Poisoning Pre-trained Models (NeurIPS 2024 · similarity 0.88)
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.
