Research map / Adversarial robustness and security / Backdoor and poisoning
Seal Your Backdoor with Variational Defense
Ivan Sabolić, Matej Grcić, Siniša Šegvić
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 |
|---|---|---|---|
| replace heuristic data splitting with latent clean-label posterior recovery (VIBE-SS-F) | backdoor defense on CIFAR-10 · CIFAR-10 (avg over 8 attacks) | ASR (%): 16.0 -> 3.2 vs best baseline ABL | Improved |
| use VIBE-SS on CIFAR-100 versus best baseline | backdoor defense on CIFAR-100 · CIFAR-100 (avg over 4 attacks) | ASR (%): 15.5 -> 0.5 vs ASD | Improved |
| add off-manifold community preprocessing (ablation) | VIBE on LC clean-label attack · CIFAR-10 LC attack | ASR (%): 14% -> 6.3% | Improved |
| add self-supervised pre-training on the poisoned data before EM (ablation) | VIBE · CIFAR-10 | ELBO / convergence | Improved |
| use full corrupted posterior p(y|x,l) instead of approximation p(y|l) (ablation) | VIBE-SS · CIFAR-10 (avg over 8 attacks) | ASR (%): 6.8 -> 3.2 | Improved |
| use frozen DINOv2 foundation-model features instead of self-supervised ResNet-18 | VIBE · CIFAR-100 transfer | ACC (%): 74.3 -> 92.3 | Improved |
| lower the poisoning rate (ablation) | VIBE vs other defenses · BadNets CIFAR-10 | ASR | No clear effect |
| replace SimCLR pre-training by SimCLR + MIMIC robust pre-training (ablation) | VIBE-SS · CIFAR-10 CTRL attack | ASR (%): 26.8 -> 4.2 | Improved |
| apply VIBE to dataset poisoned with both BadNets and LC | VIBE-SS-A vs DBD-SS · CIFAR-10 BadNets+LC | ASR (%): 99.8 -> 1.3 | Improved |
| use VIBE instead of baselines on Adap-Blend attack | VIBE-SS-F · CIFAR-10 Adap-Blend | ASR (%): 93.9 -> 9.0 vs ASD; VIBE-SS-A 36.7 | Mixed |
Closest papers by meaning
- Clean Data Can Still Carry Backdoors: Support-Persistent Backdoors for Model Reuse (NeurIPS 2026 · similarity 0.86)
- Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized Scaled Prediction Consistency (ICLR 2024 · similarity 0.86)
- Data Free Backdoor Attacks (NeurIPS 2024 · similarity 0.86)
- RPP: A Certified Poisoned-Sample Detection Framework for Backdoor Attacks under Dataset Imbalance (NeurIPS 2026 · similarity 0.85)
- SampDetox: Black-box Backdoor Defense via Perturbation-based Sample Detoxification (NeurIPS 2024 · similarity 0.85)
- Inspecting Prediction Confidence for Detecting Black-Box Backdoor Attacks (AAAI 2024 · 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.
