Abstract
Several recent studies have shown that the use of extra in-distribution data can lead to a high level of adversarial robustness. However, there is no guarantee that it will always be possible to obtain sufficient extra data for a selected dataset. In this paper, we propose a biased multi-domain adversarial training (BiaMAT) method that induces training data amplification on a primary dataset using publicly available auxiliary datasets, without requiring the class distribution match between the primary and auxiliary datasets. The proposed method can achieve increased adversarial robustness on a primary dataset by leveraging auxiliary datasets via multi-domain learning. Specifically, data amplification on both robust and non-robust features can be accomplished through the application of BiaMAT as demonstrated through a theoretical and empirical analysis. Moreover, we demonstrate that while existing methods are vulnerable to negative transfer due to the distributional discrepancy between auxiliary and primary data, the proposed method enables neural networks to flexibly leverage diverse image datasets for adversarial training by successfully handling the domain discrepancy through the application of a confidence-based selection strategy. The pre-trained models and code are available at: https://github.com/Saehyung-Lee/BiaMAT.
Introduction
Deep Neural Networks (DNNs) have achieved remarkable success in various computer vision tasks but remain notoriously fragile to adversarial examples—small, intentional perturbations that can fool a model. Adversarial Training (AT) is currently the most effective defense mechanism. However, research indicates that AT has a significantly higher sample complexity than standard training; achieving robustness requires far more data than achieving high clean accuracy.
While recent breakthroughs have leveraged massive amounts of unlabeled, in-distribution data to boost robustness, such data is not always available for specific target tasks. This research introduces BiaMAT (Biased Multi-domain Adversarial Training), a framework designed to utilize diverse, publicly available auxiliary datasets (like ImageNet or Places365) to improve robustness on a primary task (like CIFAR-10), even when the categories do not match.
Methodology: Biased Multi-domain Adversarial Training
The core of BiaMAT is a multi-domain learning approach where the primary and auxiliary tasks share a feature embedding function but utilize separate prediction heads ($f_{pri}$ and $f_{aux}$). The framework addresses the risk of "negative transfer"—where unrelated data degrades performance—through a confidence-based selection strategy.
As shown in the architecture above, auxiliary data is partitioned into two groups:
- High-Confidence Data ($X_{high}$): Images that share both "robust" (human-perceptible) and "non-robust" (imperceptible) features with the primary dataset. These are used for standard adversarial training.
- Low-Confidence Data ($X_{low}$): Images that are too different to share robust features but still contain "non-robust" features. To leverage these without confusing the model, BiaMAT trains against the expectation of random labels ($y^{ER}$). This regularizes the model against universal noise patterns without forcing it to learn incorrect semantic shapes.
Theoretical Foundation: Feature Dissociation
The effectiveness of BiaMAT rests on the dissociation of features into two categories. Robust features correspond to human-perceptible patterns (e.g., the shape of a strawberry), while non-robust features are high-frequency, imperceptible patterns that are highly predictive but easily manipulated by attackers.
BiaMAT achieves data amplification for both types. By using $X_{high}$, it expands the set of robust features the model learns. By using $X_{low}$ with random labels, it amplifies the non-robust features, teaching the model to ignore the "noise" common across diverse image domains.
Experimental Results
The authors evaluated BiaMAT against standard AT and the TRADES method across several benchmarks, using auxiliary datasets like SVHN, Places365, and ImageNet. Evaluation was conducted using AutoAttack (AA), a rigorous ensemble of adversarial attacks.
- CIFAR-10: Using ImageNet as auxiliary data, BiaMAT improved AA accuracy from 48.53% (baseline AT) to 50.78%.
- CIFAR-100: AA accuracy rose from 24.13% to 27.65% with ImageNet assistance.
- ImgNet100: Robustness increased from 29.54% to 32.88%.
Crucially, BiaMAT proved superior to traditional semi-supervised learning methods when using mismatched data. For instance, when using Places365 to assist CIFAR-10, existing methods often suffered from negative transfer, whereas BiaMAT successfully leveraged the domain gap to improve robustness.
Conclusion
BiaMAT demonstrates that adversarial robustness can be significantly enhanced by "borrowing" data from unrelated tasks. By intelligently separating auxiliary data based on confidence and employing random-label regularization for dissimilar samples, the method avoids negative transfer and achieves state-of-the-art results. Future work will focus on adaptive weighting strategies to further refine the selection process and reduce the computational overhead associated with processing large auxiliary datasets.
Summary updated