Clean Data Can Still Carry Backdoors: Support-Persistent Backdoors for Model Reuse
cs.CR, cs.LG
Submitted: 2026-10-15
Updated: 2026-10-15
Comments: Accepted to NeurIPS 2026
Project page: https://junhoo.me/backdoor
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Sources
- Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning
- Natural Backdoor Datasets
- Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization
- Breaking the Stealth-Potency Trade-off in Clean-Image Backdoors with Generative Trigger Optimization
- Poison Ink: Robust and Invisible Backdoor Attack
- Dynamic Backdoor Attacks Against Machine Learning Models
- CSF: Black-box Fingerprinting via Compositional Semantics for Text-to-Image Models
- Distilling the Knowledge in a Neural Network
- Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks
- Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks
- Deep Support Vectors
- Do not think about pink elephant!
- Backdoor Attacks Against Deep Learning Systems in the Physical World
- BadCLIP: Trigger-Aware Prompt Learning for Backdoor Attacks on CLIP
- Keeping the Bad Guys Out: Protecting and Vaccinating Deep Learning with JPEG Compression
- Countering Adversarial Images using Input Transformations
Related papers
- SoK: AI-Augmented Binary Reversing
- Relaxed Sender Anonymity for CBDC Interbank Settlement: A Zero-Knowledge Approach on Permissioned EVM
- Calibration-Family Overfit: Why Trusted Sabotage Monitors Don't Transfer Across Lineages
- Estimating the Social Cost of Corporate Data Breaches
- Efficient Fuzzy PSI under One-Sided Assumptions
- Sealing the Audit-Runtime Gap for LLM Skills