TL;DR
This paper introduces a novel black-box adversarial attack method that leverages meta-learning to train "meta adversarial perturbations" (MAPs). These MAPs act as powerful initializations for existing black-box attack techniques, significantly boosting their success rates and reducing the number of queries needed to fool deep neural networks.
What Problem Does This Paper Address?
Deep neural networks (DNNs) are increasingly used in critical applications, but they are notoriously vulnerable to adversarial examples – subtly modified inputs that cause misclassifications. While white-box attacks (where the attacker has full model knowledge) are well-studied, black-box attacks (where the attacker only has input-output access) are more practical but often suffer from low success rates and require a prohibitive number of queries to the target model. This paper aims to bridge this gap by developing a more efficient and effective black-box attack strategy.
Methods
The core innovation is the concept of Meta Adversarial Perturbations (MAPs). These are learned perturbations that are designed to be universally effective across different images and, crucially, adaptable to various target models without direct access to their internal workings.
The method involves two main stages:
- MAPs Training: MAPs are trained on surrogate models (models with known architectures and parameters) using a meta-learning approach inspired by MAML (Model-Agnostic Meta-Learning). The goal is to find a perturbation that, when applied to an input, makes it adversarial for the surrogate model, and this learned perturbation can then be quickly adapted to new, unseen data points. This training process uses a gradient-based iterative method, optimizing the perturbation to satisfy a target classification.
- Query-Efficient Black-Box Attack: Once MAPs are trained, they are used to initialize the attack on the target black-box model. The MAP is first added to the input image. If this initial perturbation is sufficient to cause a misclassification, the attack is successful with minimal queries. If not, the method employs a zero-order optimization algorithm (specifically, the RGF method) to estimate the target model's gradient and iteratively refine the adversarial example. A momentum term is incorporated into this iterative process to accelerate the attack and further reduce query counts.
Datasets
- CIFAR10: This dataset was used for both training surrogate models and evaluating the proposed attack method. It consists of RGB images across 10 classes.
- Surrogate Models: The paper used five pre-trained models (VGG13, VGG16, GoogleNet, ResNet18, ResNet34) with varying architectures and parameters as surrogate models for training MAPs.
- Victim Models: One of the pre-trained models was designated as the black-box victim model for each attack experiment.
Key Insights and Ideas
- Meta-Learning for Perturbations: Applying meta-learning to learn adversarial perturbations is a powerful idea. It allows the model to learn a general "attack strategy" that can be quickly adapted to new scenarios, rather than learning specific attacks for specific models.
- MAPs as Universal Initializers: The trained MAPs act as a "warm start" for black-box attacks. This initialization significantly reduces the need for extensive querying, as the MAP already pushes the input towards an adversarial state.
- Transferability and Adaptability: The MAPs demonstrate strong transferability and universality. They can effectively attack models they haven't seen during training, even when trained on data semantically unrelated to the attacked images, highlighting their robustness.
- Hybrid Approach Synergy: Combining the transferability of MAPs with query-based gradient estimation (enhanced by momentum) creates a highly effective hybrid attack strategy.
Key Results
- Improved Success Rates: The proposed method consistently achieved higher attack success rates compared to existing black-box attack methods like RGF, NES, and P-RGF, often by a significant margin (e.g., around 10% improvement in some cases).
- Reduced Query Counts: The method drastically reduced the number of queries required to achieve successful attacks, making it much more practical for real-world black-box scenarios. In some instances, it used fewer queries than other methods to achieve similar or better success rates.
- Invisible Distortions: The adversarial examples generated by the method had minimal visual distortions, making them difficult for humans to detect.
- Adaptability Demonstrated: Experiments showed that MAPs trained on animal images could effectively attack models with transportation images, demonstrating their adaptability to unseen data distributions.
Limitations and Future Work
- Targeted Attacks Focus: The paper primarily focuses on targeted attacks, where the goal is to misclassify an image into a specific target class. While the authors mention the method can be applied to non-targeted attacks with slight modifications, the experimental results are mainly for targeted scenarios.
- Future Work: The authors suggest exploring training more efficient MAPs on larger datasets and further boosting black-box attacks.
Summary updated