How to Train a Backdoor-Robust Model on a Poisoned Dataset without Auxiliary Data?
| Autor: | Pu, Yuwen, Chen, Jiahao, Zhou, Chunyi, Feng, Zhou, Li, Qingming, Hu, Chunqiang, Ji, Shouling |
|---|---|
| Rok vydání: | 2024 |
| Předmět: | |
| Druh dokumentu: | Working Paper |
| Popis: | Backdoor attacks have attracted wide attention from academia and industry due to their great security threat to deep neural networks (DNN). Most of the existing methods propose to conduct backdoor attacks by poisoning the training dataset with different strategies, so it's critical to identify the poisoned samples and then train a clean model on the unreliable dataset in the context of defending backdoor attacks. Although numerous backdoor countermeasure researches are proposed, their inherent weaknesses render them limited in practical scenarios, such as the requirement of enough clean samples, unstable defense performance under various attack conditions, poor defense performance against adaptive attacks, and so on.Therefore, in this paper, we are committed to overcome the above limitations and propose a more practical backdoor defense method. Concretely, we first explore the inherent relationship between the potential perturbations and the backdoor trigger, and the theoretical analysis and experimental results demonstrate that the poisoned samples perform more robustness to perturbation than the clean ones. Then, based on our key explorations, we introduce AdvrBD, an Adversarial perturbation-based and robust Backdoor Defense framework, which can effectively identify the poisoned samples and train a clean model on the poisoned dataset. Constructively, our AdvrBD eliminates the requirement for any clean samples or knowledge about the poisoned dataset (e.g., poisoning ratio), which significantly improves the practicality in real-world scenarios. Comment: 13 pages, under review |
| Databáze: | arXiv |
| Externí odkaz: |
|