Research map / Adversarial robustness and security / Backdoor and poisoning
Two Sides of the Same Coin: Learning the Backdoor to Remove the Backdoor
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a training-time backdoor defense that iteratively learns a backdoor (unlearning-based) to split poisoned from benign samples, then trains a clean model on the identified benign set, using a meta-splitting stage.
Problem
Defending neural networks against backdoors introduced via data poisoning at training time; existing methods either use a fixed loss threshold for splitting clean and poisonous samples or iteratively learn a reference model as an oracle, which can be unreliable.
Approach
Exploits the fact that models learn poisonous samples faster than benign ones: it learns a backdoored model on a split, then unlearns on a benign subset to form an iterative splitting procedure, and finally uses a learned meta-splitter to separate poisoned from clean samples before training a clean model.
Experimental setup
Baselines: ABL DBD ASD
Builds on: backdoor isolation anti-backdoor learning machine unlearning
Closest papers by meaning
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.91)
- Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks (ICLR 2025 · similarity 0.90)
- Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized Scaled Prediction Consistency (ICLR 2024 · similarity 0.89)
- SampDetox: Black-box Backdoor Defense via Perturbation-based Sample Detoxification (NeurIPS 2024 · similarity 0.89)
- Mitigating Backdoor Attack by Injecting Proactive Defensive Backdoor (NeurIPS 2024 · similarity 0.89)
- Wicked Oddities: Selectively Poisoning for Effective Clean-Label Backdoor Attacks (ICLR 2025 · similarity 0.89)
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.