TOP: Backdoor Detection in Neural Networks via Transferability of Perturbation
arXiv:2103.10274
Abstract
Deep neural networks (DNNs) are vulnerable to "backdoor" poisoning attacks, in which an adversary implants a secret trigger into an otherwise normally functioning model. Detection of backdoors in trained models without access to the training data or example triggers is an important open problem. In this paper, we identify an interesting property of these models: adversarial perturbations transfer from image to image more readily in poisoned models than in clean models. This holds for a variety of model and trigger types, including triggers that are not linearly separable from clean data. We use this feature to detect poisoned models in the TrojAI benchmark, as well as additional models.
References in corpus (5)
- Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning
- advertorch v0.1: An Adversarial Robustness Toolbox based on PyTorch
- Backdoor Attacks and Countermeasures on Deep Learning: A Comprehensive Review
- Practical Detection of Trojan Neural Networks: Data-Limited and Data-Free Cases
- Baseline Pruning-Based Approach to Trojan Detection in Neural Networks