Boosting Adversarial Attacks with Momentum
By Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, Jianguo Li. 2017
3,379 citations recorded by the source corpus.
Deep neural networks are vulnerable to adversarial examples, which poses security concerns on these algorithms due to the potentially severe consequences. Adversarial attacks serve as an important surrogate to evaluate the robustness of deep learning models before they are deployed. However, most of existing adversarial attacks can only fool a black-box model with a low success rate. To address this issue, we propose a broad class of momentum-based iterative algorithms to boost adversarial attacks. By integrating the momentum term into the iterative process for attacks, our methods can stabilize update directions and escape from poor local maxima during the iterations, resulting in more transferable adversarial examples. To further improve the success rates for black-box attacks, we apply momentum iterative algorithms to an ensemble of models, and show that the adversarially trained models with a strong defense ability are also vulnerable to our black-box attacks. We hope that the proposed methods will serve as a benchmark for evaluating the robustness of various deep models and defense methods. With this method, we won the first places in NIPS 2017 Non-targeted Adversarial Attack and Targeted Adversarial Attack competitions.
Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Safety, Robustness & Interpretability
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Yoshua Bengio
- Bernt Schiele
- Quoc V. Le
- Percy Liang
- Jun Zhu
- Mohit Bansal
- Alan L. Yuille
- Sergey Levine
In Computer Vision and Pattern Recognition
- Luc Van Gool
- Dacheng Tao
- Alan L. Yuille
- Ming-Hsuan Yang
- Bernt Schiele
- Xiaogang Wang
- Trevor Darrell
- Andrea Vedaldi
In Safety, Robustness & Interpretability
- Matthias Hein
- Nicholas Carlini
- Tianyu Pang
- Jun Zhu
- Yinpeng Dong
- Florian Tramèr
- Nicolas Papernot
- Hang Su
Citations in the corpus
- Adversarial Example Games (2020, methodology): “Throughout our experiments we rely on four representative blackbox attack strategies adapted to the NoBox setting: the Momentum-Iterative Attack (MI-Attack) [21], the Input Diversity (DI-Attack)[70], the Translation-Invariant (TID-Attack) [22] and the Skip Gradient Method (SGM-Attack) [68].”
- Feature Denoising for Improving Adversarial Robustness (2018, background): “4We have also evaluated other attackers, including FGSM [6], iterative FGSM [12], and its momentum variant [3].”
- Ensemble Adversarial Training: Attacks and Defenses (2017, background): “By incorporating additional optimization techniques (e.g., momentum (Dong et al., 2018), data-augmentation (Xie et al., 2019b; Dong et al., 2019) or gradients from a residual network’s “skip connections” (Wu et al., 2020)), these attacks substantially improve the transferability of adversarial…”
- Adversarial Examples: Attacks and Defenses for Deep Learning (2017, methodology): “[73] applied momentum to FGSM to generate adversarial examples more iteratively.”
- Towards the first adversarially robust neural network model on MNIST (2018, methodology): “Other ideas in similar directions include the use of denoising autoencoders in (Gu & Rigazio, 2014) and (Liao et al., 2017) as well as MagNets (Meng & Chen, 2017) (which projects or rejects inputs depending on their distance to the data manifold).”
- Prior Networks for Detection of Adversarial Attacks (2018, methodology): “This work investigates the detection of Fast Gradient Sign Method (FGSM) [11], Basic Iterative Method (BIM) [13] and Momentum Iterative Method (MIM) [16] adversarial attacks using measures of model or distributional uncertainty derived from either a Monte-Carlo dropout derived ensemble or Prior Netw”
- Theoretically Principled Trade-off between Robustness and Accuracy (2019, methodology): “Other attack methods include MI-FGSM (Dong et al., 2018) and LBFGS attacks (Tabacof & Valle, 2016).”
- Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey (2018, methodology): “[168] proposed an iterative method to boost adversarial attacks for black-box scenarios.”
- Technical Report on the CleverHans v2.1.0 Adversarial Examples Library (2016, methodology): “The momentum iterative method (MIM) was introduced by Dong et al. [10].”
- Fast is better than free: Revisiting adversarial training (2020, methodology): “…adversarial training defense have been augmented with various techniques, such as optimization tricks like momentum to improve the adversary (Dong et al., 2018), combination with other heuristic defenses like matrix estimation (Yang et al., 2019) or logit pairing (Mosbach et al., 2018), and…”
- Disentangling Adversarial Robustness and Generalization (2018, background): “Adversarial attacks on deep networks range from white-box attacks [97, 30, 49, 71, 66, 61, 14, 78, 21, 59], with full access to the model (weights, gradients etc.”
- On Adaptive Attacks to Adversarial Example Defenses (2020, background): “However, BIM (Kurakin et al., 2016), PGD (Madry et al., 2017) and MIM (Dong et al., 2018) are extremely similar and if one of these attacks fails the others are likely to fail as well.”
- ComDefend: An Efficient Image Compression Model to Defend Adversarial Examples (2018, methodology): “We see that the proposed method achieves the higher defensive accuracy against FGSM, DeepFool, MI-FGSM [5] and C&W, and the competitive accuracy against IFGSM compared with HGD, which demonstrates the effectiveness of the proposed method.”
- Universal Adversarial Training (2018, methodology): “iterative optimization procedures [8, 16, 2].”
- advertorch v0.1: An Adversarial Robustness Toolbox based on PyTorch (2019, background): “…categories: generating strong and efficient attacks (Goodfellow et al., 2014; Carlini and Wagner, 2017; Brendel et al., 2017; Xiao et al., 2018; Dong et al., 2018); detecting adversarial examples (Metzen et al., 2017; Ma et al., 2018; Feinman et al., 2017); defending already trained models…”
- Adversarially Robust Few-Shot Learning: A Meta-Learning Approach (2019, methodology): “ersarially queried R2-D2 model to a variety of powerful adversarial attacks. We implement the momentum iterated fast gradient sign method (MI-FGSM), DeepFool, and 20-step PGD with 20 random restarts (Dong et al., 2018; Moosavi-Dezfooliet al., 2016; Madry et al., 2017). Our adversarially queried mode”
- Regional Homogeneity: Towards Learning Transferable Universal Adversarial Perturbations Against Defenses (2019, methodology): “[14], Zhou et al.”