Research map / Adversarial robustness and security / Attacks on text and graphs
Hidden Poison: Machine Unlearning Enables Camouflaged Poisoning Attacks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 gradient-matched poison and camouflage points, then remove the camouflage points through unlearning | Camouflaged poisoning attack · CIFAR-10, Imagenette, Imagewoof | targeted attack success: Camouflaging succeeded at least 70% of the time for VGG-11, VGG-16, ResNet-18, and ResNet-34; 100% on Imagenette with VG | Helped |
| Use gradient-matching rather than label-flipping camouflage (ablation) | Linear SVM camouflaged poisoning · Binary-CIFAR-10 | camouflaging success: 71.5% vs 100% at ε=8; 40% vs 70% at ε=16 | Mixed |
| Increase camouflage budget from 0.5% to 1% with poison budget fixed at 1% (ablation) | Camouflaged poisoning on ResNet-18 · CIFAR-10 | camouflaging success: 40% -> 80% | Helped |
| Increase camouflage budget from 1% to 1.5% with poison budget fixed at 1% (ablation) | Camouflaged poisoning on ResNet-18 · CIFAR-10 | camouflaging success: 80% -> 80% | No clear effect |
| Increase poison budget from 1% to 1.5% while keeping camouflage budget at 1% (ablation) | Camouflaged poisoning on ResNet-18 · CIFAR-10 | camouflaging success: 80% -> 20% | Hurt |
| Randomly delete poison and camouflage samples (ablation) | Camouflaged poisoning on ResNet-18 · CIFAR-10 | camouflaging success: 75%, 33%, 100%, 80%, 100% with 5%, 10%, 20%, 30%, 40% removed | Mixed |
| Add horizontal-flip augmentation (ablation) | Camouflaged poisoning on ResNet-18 · CIFAR-10 | camouflaging success: 20% -> 33% | Helped |
| Add horizontal flips and random rotations (ablation) | Camouflaged poisoning on ResNet-18 · CIFAR-10 | camouflaging success: 20% -> 100% | Helped |
| Add data augmentation during training and retraining (ablation) | Camouflaged poisoning on ResNet-18 · CIFAR-10 | poisoning success: 100% -> 86% -> 60% | Hurt |
Closest papers by meaning
- Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025 · similarity 0.92)
- Backdoor Attacks via Machine Unlearning (AAAI 2024 · similarity 0.90)
- Unlearn and Burn: Adversarial Machine Unlearning Requests Destroy Model Accuracy (ICLR 2025 · similarity 0.90)
- Adversarial Machine Unlearning (ICLR 2025 · similarity 0.89)
- Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable (NeurIPS 2024 · similarity 0.88)
- Universal Backdoor Attacks (ICLR 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.