Research map / Continual learning / Online and replay-based
Poisoning Generative Replay in Continual Learning to Promote Forgetting
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 input-aware dirty-label backdoor poisoning (CIAP) to replayer training data | generative replay continual learning · split-MNIST | test accuracy on past tasks: past-task accuracy drops to around 25% from replay baseline | Helped |
| Add input-aware backdoor poisoning while keeping current-task accuracy | generative replay continual learning · split-MNIST | test accuracy on current task: current-task accuracy stays near 100%, same as clean | No clear effect |
| Reduce poison ratio to 5% backdoor and 5% cross (ablation) | CIAP · split-MNIST | test accuracy on past tasks: attack weaker: past accuracy above 70% vs about 25% at 25%/15% | Hurt |
| Use a surrogate classifier with different architecture than victim (5%/5% poisons) (ablation) | CIAP · split-MNIST | test accuracy on past tasks: past-task accuracy reduced from 90% (clean) to 35% | Helped |
| Apply nu-SVM outlier filter to replayed samples (ablation) | CIAP · split-MNIST | test accuracy on past tasks: filtering helps only a little; past accuracy still below 50% | No clear effect |
| Apply nu-SVM outlier filter to replayed samples (ablation) | CIAP · split-CIFAR-10 | test accuracy on past tasks: drops to almost 0 across nu values | No clear effect |
| Replace input-aware trigger with static white-square backdoor (ablation) | replayer trigger preservation · split-MNIST | attack success rate after replay: static ASR 100% vs IAB ASR 0 (i.e., IAB is discarded) | Hurt |
| Use input-aware (per-example) trigger rather than static trigger (ablation) | cWGAN replayer · split-MNIST | fraction of backdoored images produced by replayer: IAB yields 40% backdoored images at 90% poisoned training vs static retaining nearly all | Helped |
| Replace clean replay with poisoned replay via CIAP on FashionMNIST-MNIST (cWGAN) | generative replay · FashionMNIST-MNIST | test accuracy on earlier task: earlier-task accuracy 17% under attack vs 66% clean replay | Helped |
| Replace cWGAN with cVAE replayer under CIAP attack on FashionMNIST-MNIST | generative replay · FashionMNIST-MNIST | test accuracy on earlier task: earlier-task accuracy drops from 91% to 27% | Helped |
| Apply CIAP poisoning to split-CIFAR-10 replay | generative replay · split-CIFAR-10 | test accuracy on past tasks: past-task accuracy drops to 0% after poisoning, erasing DGR gains | Helped |
| Naive label flipping (25%) instead of input-aware poisoning (ablation) | naive poisoning baseline · split-MNIST | test accuracy on current task: current-task accuracy drops below 20% from task 2, failing stealth | Hurt |
Closest papers by meaning
- BrainWash: A Poisoning Attack to Forget in Continual Learning (CVPR 2024 · similarity 0.89)
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.87)
- Attack on Prompt: Backdoor Attack in Prompt-Based Continual Learning (AAAI 2025 · similarity 0.87)
- Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning (ICML 2024 · similarity 0.86)
- Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization (ICML 2024 · similarity 0.85)
- ZeroFlow: Overcoming Catastrophic Forgetting is Easier than You Think (ICML 2025 · similarity 0.85)
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.