Practical Black-Box Attacks against Machine Learning
By Nicolas Papernot, Patrick Drew McDaniel, Ian J. Goodfellow, Somesh Jha, Z. Berkay Celik, Ananthram Swami. 2016
4,113 citations recorded by the source corpus.
Machine learning (ML) models, e.g., deep neural networks (DNNs), are vulnerable to adversarial examples: malicious inputs modified to yield erroneous model outputs, while appearing unmodified to human observers. Potential attacks include having malicious content like malware identified as legitimate or controlling vehicle behavior. Yet, all existing adversarial example attacks require knowledge of either the model internals or its training data. We introduce the first practical demonstration of an attacker controlling a remotely hosted DNN with no such knowledge. Indeed, the only capability of our black-box adversary is to observe labels given by the DNN to chosen inputs. Our attack strategy consists in training a local model to substitute for the target DNN, using inputs synthetically generated by an adversary and labeled by the target DNN. We use the local substitute to craft adversarial examples, and find that they are misclassified by the targeted DNN. To perform a real-world and properly-blinded evaluation, we attack a DNN hosted by MetaMind, an online deep learning API. We find that their DNN misclassifies 84.24% of the adversarial examples crafted with our substitute. We demonstrate the general applicability of our strategy to many ML techniques by conducting the same attack against models hosted by Amazon and Google, using logistic regression substitutes. They yield adversarial examples misclassified by Amazon and Google at rates of 96.19% and 88.94%. We also find that this black-box attack strategy is capable of evading defense strategies previously found to make adversarial example crafting harder.
Subfields: Artificial Intelligence, Safety, Robustness & Interpretability, Signal Processing
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
- Percy Liang
- Mohit Bansal
- Yoshua Bengio
- Quoc V. Le
- Bernt Schiele
- Jianfeng Gao
- Sergey Levine
- Luke S. Zettlemoyer
In Safety, Robustness & Interpretability
- Nicholas Carlini
- Matthias Hein
- Nicolas Papernot
- Bo Li
- Cho-Jui Hsieh
- Bernt Schiele
- Huan Zhang
- Florian Tramèr
In Signal Processing
- Nicholas Carlini
- Bo Li
- Nicolas Papernot
- Aleksander Madry
- Battista Biggio
- Matthias Hein
- Bernt Schiele
- Dawn Song
Citations in the corpus
- Adversarial Example Games (2020, background): “This setting is more permissive than the semi-whtiebox and the blackbox setting, the latter of which the attacker has only access to the prediction [54, 55] or sometimes the predicted confidence [33].”
- Analysis Methods in Neural Language Processing: A Survey (2018, background): “Adversarial examples can be generated using access to model parameters, also known as white-box attacks, or without such access, with black-box attacks (Papernot et al., 2016a, 2017; Narodyt-ska and Kasiviswanathan, 2017; Liu et al., 2017).”
- Stress Test Evaluation for Natural Language Inference (2018, background): “However, machine learning models are known to exploit idiosyncrasies of how the data was produced, allowing them to imitate desired behavior using tricks such as pattern matching (Levesque, 2014; Rimell et al., 2009; Papernot et al., 2017).”
- Synthetic and Natural Noise Both Break Neural Machine Translation (2017, background): “…important distinction is often drawn between white-box attacks, where adversarial examples are generated with access to the model parameters, and black-box attacks, where examples are generated without such access (Papernot et al., 2016; 2017; Liu et al., 2017; Narodytska & Kasiviswanathan, 2017).”
- How To Backdoor Federated Learning (2018, background): “Adversarial examples [21, 37, 52] are deliberately crafted to be misclassified by the model.”
- Wild Patterns: Ten Years After the Rise of Adversarial Machine Learning (2017, methodology): “Although the problem of learning a surrogate model while minimizing the number of queries can be casted as an active learning problem, to our knowledge well-established active learning algorithms have not yet been compared against such recently-proposed approaches [14].”
- Adversarial Training and Robustness for Multiple Perturbations (2019, background): “This is an indication of gradient-masking [27, 40, 1].”
- Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning (2017, background): “In particular, a body of literature studying adversarial examples belongs to this category [29], [53], [36], [48], [54], [52], [15], [40], [25].”
- Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models (2017, methodology): “The most relevant prior work is a variant of transfer attacks in which the training set needed to learn the substitute model is replaced by a synthetic dataset (Papernot et al., 2017b).”
- A Rotation and a Translation Suffice: Fooling CNNs with Simple Transformations (2017, methodology): “There is a long line of work (Szegedy et al., 2013; Goodfellow et al., 2014; Kurakin et al., 2016a;b; Sharif et al., 2016; Moosavi-Dezfooli et al., 2016; Carlini & Wagner, 2016; Papernot et al., 2017; Madry et al., 2017; Athalye et al., 2017) on methods for constructing adversarial perturbations Equ”
- BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain (2017, background): “Follow-on work improved the speed at which adversarial examples could be created [33], demonstrated that adversarial examples could be found even if only black-box access to the target model was available [34], and even discovered universal adversarial perturbations [35] that could cause different i”
- ZOO: Zeroth Order Optimization Based Black-box Attacks to Deep Neural Networks without Training Substitute Models (2017, methodology): “Experimental results in Section 4 show that our attack significantly outperforms [35], which can be explained by the fact that our attack inherits the effectiveness of the state-of-the-art C&W attack, and also by the fact that zeroth order optimization allows direct adversarial attacks to a targeted”
- Ensemble Adversarial Training: Attacks and Defenses (2017, methodology): “[13] show that a model trained on perturbations with small norm is vulnerable to larger transferred perturbations.”