TL;DR
This paper introduces PUBDEF, a novel defense against adversarial attacks that leverages publicly available pre-trained models as surrogates. Unlike traditional defenses that assume a strong white-box threat model, PUBDEF focuses on a more realistic scenario where attackers exploit readily accessible models. PUBDEF achieves significant adversarial robustness with minimal impact on clean accuracy, outperforming existing methods.
What Problem Does This Paper Address?
The field of adversarial machine learning has long grappled with defending models against adversarial examples – subtly perturbed inputs that cause misclassifications. A dominant threat model in research assumes "white-box" access, where attackers know everything about the target model. However, this is often unrealistic. A more practical threat is "transfer attacks," where an attacker uses a different, publicly available model (a surrogate) to craft an attack that then transfers to the target model. This is particularly concerning because public model repositories are vast, making it easy for attackers to find suitable surrogates. Existing defenses often suffer from a trade-off: improving adversarial robustness significantly degrades clean accuracy, making them impractical for real-world deployment. This paper addresses this gap by proposing a defense that is robust to transfer attacks from public models while maintaining high clean accuracy.
Methods
The core of PUBDEF is a defense strategy that trains a model to be robust against transfer attacks originating from a diverse set of publicly available "source" models. The authors frame this as a game-theoretic problem, where the attacker chooses a source model and an attack algorithm, and the defender aims to minimize the impact of these attacks. PUBDEF employs a specialized training process that incorporates these transfer attacks. A key aspect is the selection of source models: instead of training against all available models (computationally prohibitive), they propose a heuristic to select a diverse subset, ensuring representation from different training methodologies (normal, l∞-robust, l2-robust, corruption-robust). This approach aims to generalize well to unseen attacks and source models.
Datasets
The authors evaluate their method on three standard image classification datasets:
- CIFAR-10
- CIFAR-100
- ImageNet
Key Insights and Ideas
- Realistic Threat Model: The paper convincingly argues that transfer attacks from public models (TAPM) represent a more prevalent and practical threat than the traditional white-box model.
- PUBDEF's Generalization: PUBDEF demonstrates remarkable generalization capabilities, maintaining robustness even against transfer attacks from source models and algorithms it was not explicitly trained on. This suggests that adversarial examples might occupy a low-dimensional subspace, making them learnable even with limited training data.
- Minimal Clean Accuracy Drop: A significant advantage of PUBDEF is its ability to achieve high adversarial accuracy with a negligible drop in clean accuracy, a critical factor for practical deployment.
- Source Model Diversity is Key: The heuristic for selecting diverse source models (representing different training paradigms) is crucial for PUBDEF's effectiveness, highlighting that robustness against one type of attack doesn't automatically imply robustness against others.
Key Results
- On ImageNet, PUBDEF achieves 62% accuracy against the strongest transfer attack, significantly outperforming the best adversarially trained model (36%).
- PUBDEF's clean accuracy on ImageNet is only 2% lower than an undefended model (78% vs. 80%).
- For CIFAR-10, PUBDEF achieves 88.6% adversarial accuracy, a substantial improvement over previous defenses.
- The defense shows strong generalization, with only a small drop in adversarial accuracy (1.7-7.6%) against unseen attacks.
Limitations and Future Work
The authors acknowledge several limitations:
- White-box Vulnerability: PUBDEF is not robust to white-box attacks. If model weights are not kept secret, attackers can reverse-engineer them and mount white-box attacks.
- Sophisticated Attackers: Highly resourced attackers might still be able to bypass the defense by training their own private surrogate models.
- Lp-Norm Bounded Attacks: The defense is specifically designed for lp-norm bounded attacks and does not cover broader attack strategies.
Future work could explore combining PUBDEF with other defense mechanisms, investigating more sophisticated ensemble-based attacks, and extending the defense to other types of attacks beyond lp-norm bounds.
Summary updated