Universal Litmus Patterns: Revealing Backdoor Attacks in CNNs
Soheil Kolouri, Aniruddha Saha, Hamed Pirsiavash, Heiko Hoffmann
摘要
The unprecedented success of deep neural networks in many applications has made these networks a prime target for adversarial exploitation. In this paper, we introduce a benchmark technique for detecting backdoor attacks (aka Trojan attacks) on deep convolutional neural networks (CNNs). We introduce the concept of Universal Litmus Patterns (ULPs), which enable one to reveal backdoor attacks by feeding these universal patterns to the network and analyzing the output (i.e., classifying the network as 'clean' or 'corrupted'). This detection is fast because it requires only a few forward passes through a CNN. We demonstrate the effectiveness of ULPs for detecting backdoor attacks on thousands of networks with different architectures trained on four benchmark datasets, namely the German Traffic Sign Recognition Benchmark (GTSRB), MNIST, CIFAR10, and Tiny-ImageNet. The codes and train/test models for this paper can be found here: https://umbcvision . github.io/Universal-Litmus-Patterns/.
问 Lune
几分钟读透这篇 Paper,不用再花一下午。
Lune 能读完全文,也能读引用它的顶会 Paper,讲清它的结论、证据和局限,每一点都注明出自哪篇 Paper。不用硬啃,也不靠猜。
已有 77 篇顶会 Paper 引用了它。别等审稿人问起,先弄清它们发现了什么。
脉络
- Backdoor Scanning for Deep Neural Networks through K-Arm Optimization 引用了 Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks
- Backdoor Scanning for Deep Neural Networks through K-Arm Optimization 引用了 Trojaning Attack on Neural Networks
- Backdoor Scanning for Deep Neural Networks through K-Arm Optimization 引用了 Deep Feature Space Trojan Attack of Neural Networks by Controlled Detoxification
- Deep Feature Space Trojan Attack of Neural Networks by Controlled Detoxification 引用了 Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks
- Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks 引用了 Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks
- Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks 引用了 Trojaning Attack on Neural Networks
- Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks 引用了 Towards Evaluating the Robustness of Neural Networks
- When Does Machine Learning FAIL? Generalized Transferability for Evasion and Poisoning Attacks 引用了 Trojaning Attack on Neural Networks
- When Does Machine Learning FAIL? Generalized Transferability for Evasion and Poisoning Attacks 引用了 Towards Evaluating the Robustness of Neural Networks
- Black-box Detection of Backdoor Attacks with Limited Information and Data 引用了 Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks
- Black-box Detection of Backdoor Attacks with Limited Information and Data 引用了 Trojaning Attack on Neural Networks
- Black-box Detection of Backdoor Attacks with Limited Information and Data 引用了 Towards Evaluating the Robustness of Neural Networks
- Anti-Backdoor Learning: Training Clean Models on Poisoned Data 引用了 Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks
- Anti-Backdoor Learning: Training Clean Models on Poisoned Data 引用了 Trojaning Attack on Neural Networks
- Anti-Backdoor Learning: Training Clean Models on Poisoned Data 引用了 Backdoor Scanning for Deep Neural Networks through K-Arm Optimization
- Anti-Backdoor Learning: Training Clean Models on Poisoned Data 引用了 Deep Feature Space Trojan Attack of Neural Networks by Controlled Detoxification
- Anti-Backdoor Learning: Training Clean Models on Poisoned Data 引用了 Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks
引用它的顶会 Paper77
- Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural NetworksYige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu 等ICLR 2021 · 548 次引用
- Anti-Backdoor Learning: Training Clean Models on Poisoned DataYige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu 等NeurIPS 2021 · 503 次引用
- Deep Feature Space Trojan Attack of Neural Networks by Controlled DetoxificationSiyuan Cheng, Yingqi Liu, Shiqing Ma, Xiangyu ZhangAAAI 2021 · 191 次引用
- Backdoor Scanning for Deep Neural Networks through K-Arm OptimizationGuangyu Shen, Yingqi Liu, Guanhong Tao, Shengwei An 等ICML 2021 · 137 次引用
- Black-box Detection of Backdoor Attacks with Limited Information and DataYinpeng Dong, Xiao Yang, Zhijie Deng, Tianyu Pang 等ICCV 2021 · 128 次引用
它引用的 Paper4
- Towards Evaluating the Robustness of Neural NetworksNicholas Carlini, David A. WagnerS&P 2017 · 9,786 次引用
- Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural NetworksBolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li 等S&P 2019 · 1,801 次引用
- Trojaning Attack on Neural NetworksYingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee 等NDSS 2018 · 1,377 次引用
- When Does Machine Learning FAIL? Generalized Transferability for Evasion and Poisoning AttacksOctavian Suciu, Radu Marginean, Yigitcan Kaya, Hal Daumé III 等USENIX Security 2018 · 321 次引用
相关 Paper
- Beating Backdoor Attack at Its Own GameMin Liu, Alberto L. Sangiovanni-Vincentelli, Xiangyu YueICCV 2023 · 19 次引用
- AdvDoor: adversarial backdoor attack of deep learning systemQuan Zhang, Yifeng Ding, Yongqiang Tian, Jianmin Guo 等ISSTA 2021 · 57 次引用
- UMD: Unsupervised Model Detection for X2X Backdoor AttacksZhen Xiang, Zidi Xiong, Bo LiICML 2023 · 27 次引用
- FreeEagle: Detecting Complex Neural Trojans in Data-Free CasesChong Fu, Xuhong Zhang, Shouling Ji, Ting Wang 等USENIX Security 2023
- ABS: Scanning Neural Networks for Back-doors by Artificial Brain StimulationYingqi Liu, Wen-Chuan Lee, Guanhong Tao, Shiqing Ma 等CCS 2019 · 531 次引用
