Research map / Adversarial robustness and security / Backdoor and poisoning
Sleeper Agent: Scalable Hidden Trigger Backdoors for Neural Networks Trained from Scratch
Hossein Souri, Liam H Fowl, Rama Chellappa, Micah Goldblum, Tom Goldstein
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 Sleeper Agent hidden-trigger poisoning | ResNet-18 trained from scratch · CIFAR-10 | Attack Success Rate (%): 3.50 \% -> 78.84\% | Improved |
| Use Sleeper Agent hidden-trigger poisoning | MobileNetV2 trained from scratch · CIFAR-10 | Attack Success Rate (%): 3.76 \% -> 75.96\% | Improved |
| Use Sleeper Agent hidden-trigger poisoning | VGG11 trained from scratch · CIFAR-10 | Attack Success Rate (%): 5.02 \% -> 86.60\% | Improved |
| Increase poison budget from 0.1% to 1% | Sleeper Agent on ResNet-18 · CIFAR-10 | Attack Success Rate (%): 24.71\ (\pm 4.10) -> 85.27\ (\pm 5.90) | Improved |
| Randomize patch location rather than fixing it (ablation) | Sleeper Agent on ResNet-18 · CIFAR-10 | Attack Success Rate (%): 19.25\ (\pm 3.01) -> 33.95 \ (\pm 4.57) | Improved |
| Add surrogate retraining during poison crafting (ablation) | Sleeper Agent with random patch location · CIFAR-10 | Attack Success Rate (%): 33.95 \ (\pm 4.57) -> 59.42 \ (\pm 5.78) | Improved |
| Select poisons by maximum gradient norm (ablation) | Sleeper Agent with random patch location · CIFAR-10 | Attack Success Rate (%): 33.95 \ (\pm 4.57) -> 63.49 \ (\pm 6.13) | Improved |
| Combine retraining and gradient-norm data selection (ablation) | Sleeper Agent with random patch location · CIFAR-10 | Attack Success Rate (%): 33.95 \ (\pm 4.57) -> 85.27 \ (\pm 5.90) | Improved |
| Retrain the surrogate twice during crafting (ablation) | Sleeper Agent · ImageNet | Attack Success Rate (%): 22.00 \ (\pm 5.65) -> 44.00 \ (\pm 6.73) | Improved |
| Increase retraining factor from T=4 to T=8 (ablation) | Sleeper Agent on ResNet-18 · CIFAR-10 | Attack Success Rate (%): 85.27 \ (\pm 5.90) -> 86.48 \ (\pm 6.26) | No clear effect |
| Increase the perturbation bound from 8/255 to 16/255 | Sleeper Agent on ResNet-18 · CIFAR-10 | Attack Success Rate (%): 37.32 \ (\pm 8.33) -> 85.27 \ (\pm 5.90) | Improved |
| Increase same-architecture ensemble size from S=1 to S=4 | Sleeper Agent without retraining · CIFAR-10 | Attack Success Rate (%): 63.49\ (\pm 6.13) -> 74.81 \ (\pm 4.10) | Improved |
Later papers that compared with this method
4 later accepted papers report Sleeper Agent as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.
- Injection, Attack and Erasure: Revocable Backdoor Attacks via Machine Unlearning (AAAI 2026 · better on 4 of 4 shared results)
- Sharpness-Aware Data Poisoning Attack (ICLR 2024 · better on 3 of 3 shared results)
- Generalization Bound and New Algorithm for Clean-Label Backdoor Attack (ICML 2024 · better on 1 of 1 shared result)
- Untargeted Backdoor Watermark: Towards Harmless and Stealthy Dataset Copyright Protection (NeurIPS 2022 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Hidden Trigger Backdoor Attacks (AAAI 2020 · similarity 0.90)
- SleeperNets: Universal Backdoor Poisoning Attacks Against Reinforcement Learning Agents (NeurIPS 2024 · similarity 0.88)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.87)
- From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks (AAAI 2024 · similarity 0.87)
- Trapping Attacker in Dilemma: Examining Internal Correlations and External Influences of Trigger for Defending GNN Backdoors (NeurIPS 2026 · similarity 0.86)
- LOTUS: Evasive and Resilient Backdoor Attacks through Sub-Partitioning (CVPR 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.
