TL;DR
This paper introduces AutoAttack, a novel, parameter-free ensemble of diverse adversarial attacks designed to reliably evaluate the robustness of machine learning models. By combining improved versions of PGD with existing attacks, AutoAttack successfully breaks over 50 recently proposed defenses, often revealing significantly lower robust accuracy than previously reported, highlighting widespread issues in adversarial robustness evaluation.
What Problem Does This Paper Address?
The field of adversarial robustness in machine learning has seen a surge in defense mechanisms, but progress is significantly hampered by insufficient and often flawed evaluation methods. Many proposed defenses appear robust only because the attacks used to test them are not strong enough or are improperly configured. Common pitfalls include suboptimal hyperparameter tuning of attacks (like PGD's step size), gradient masking (where gradients become uninformative), and a lack of diversity in the attacks employed. This makes it difficult to identify truly robust models and hinders the development of reliable defenses, especially for safety-critical applications. The authors aim to provide a standardized, automated, and robust evaluation protocol that overcomes these limitations.
Methods
The core contribution is the AutoAttack ensemble, which combines four distinct attacks:
- Auto-PGD (APGD): The authors propose two extensions to the standard Projected Gradient Descent (PGD) attack.
- Auto-PGD (APGDCE): This variant addresses the issue of suboptimal step sizes by introducing an adaptive, budget-aware step size selection mechanism. It dynamically adjusts the step size based on the optimization progress and restarts from the best-found point when the step size is reduced. This makes it parameter-free beyond the number of iterations.
- APGDDLR: This variant uses a novel Difference of Logits Ratio (DLR) loss function. Unlike the standard cross-entropy loss, DLR is invariant to scaling and shifting of the classifier's logits, making it more resistant to gradient masking and providing a more direct interpretation of the classifier's decision boundary.
- FAB-Attack (FABT): A white-box attack that aims to find the minimal perturbation needed for misclassification. The targeted version (FABT) is used for better scalability.
- Square Attack: A score-based black-box attack that relies on random search and does not require gradient information, making it effective against defenses that employ gradient masking.
AutoAttack combines these attacks, running each with fixed, non-tuned hyperparameters, to provide a worst-case evaluation of robustness.
Datasets
The paper evaluates its proposed methods on several standard datasets commonly used in adversarial robustness research:
- MNIST
- CIFAR-10
- CIFAR-100
- ImageNet
The authors test their ensemble on models from papers published at top machine learning and computer vision venues, covering a wide range of architectures and defense strategies.
Key Insights and Ideas
- The "state-of-the-art" is often an illusion: Many defenses that appear robust are easily broken by stronger, well-tuned, or diverse attacks, indicating a systemic issue in evaluation practices.
- PGD's fixed step size is a major weakness: The standard PGD attack's performance is highly sensitive to its step size, which is often suboptimally chosen, leading to an overestimation of robustness. Auto-PGD's adaptive step size mechanism directly addresses this.
- Cross-entropy loss is susceptible to gradient masking: The cross-entropy loss's lack of invariance to logit scaling can be exploited by defenses to create "gradient masking," making attacks ineffective. The proposed DLR loss is more robust to this.
- Ensembling diverse attacks is crucial: No single attack is sufficient. Combining white-box (APGD, FABT) and black-box (Square Attack) methods, each with different strengths, provides a more comprehensive and reliable robustness assessment.
- Parameter-free evaluation is achievable and necessary: The goal is to create an evaluation protocol that doesn't require manual tuning for each new defense, making it accessible and reproducible.
Key Results
- AutoAttack breaks over 50 defenses: The ensemble was applied to models from 35 recent papers, and in all but one case, AutoAttack achieved lower robust test accuracy than reported by the original authors.
- Significant robustness overestimation: In 13 out of 49 cases, AutoAttack reduced robust accuracy by more than 10%, and in 8 cases by more than 30%, revealing previously undetected vulnerabilities.
- Auto-PGD outperforms standard PGD: Both APGD variants consistently achieved better (lower) robust accuracy and higher (better) loss values compared to PGD with fixed step sizes and momentum, across various losses and datasets.
- DLR loss shows promise: The DLR loss generally leads to more stable and lower robust accuracy results compared to cross-entropy, especially in resisting gradient masking.
- AutoAttack is computationally affordable: Despite its effectiveness, the ensemble's computational cost is comparable to existing evaluation methods, making it practical.
Limitations and Future Work
- Discontinuous classifiers: The paper notes that APGDCE and APGDDLR can still overestimate robustness for highly discontinuous classifier functions (e.g., the defense by Xiao et al., 2020). However, the black-box Square Attack was still effective in these cases, highlighting the value of the ensemble.
- Randomized defenses: While the paper adapts AutoAttack for randomized defenses by averaging results over multiple runs, it acknowledges that further research might be needed for more robust evaluation in this area.
- Future work: The authors suggest that AutoAttack should become a standard minimal test for any new defense. They also imply that further exploration of novel loss functions and attack strategies could continue to improve robustness evaluation.
Summary updated