Adversarial attacks are often easy to generate when the attacker has full access to a model's weights (the "white-box" setting). However, these attacks frequently fail when applied to a different, unknown "black-box" model. This failure occurs because the generated perturbations "overfit" to the specific decision boundaries and gradient quirks of the source model.
To bridge this gap, a growing body of research focuses on transferability: creating adversarial examples on a surrogate model that remain effective against diverse, defended, and even structurally different target models. The core insight across this field is that transferability is not a single trick, but a modular optimization problem. By stacking gradient stabilizers, input transformations, and adaptive ensembles, researchers have pushed black-box success rates from mediocre to nearly 95%.
The Overfitting Problem in Adversarial Space
Standard iterative attacks, like the Basic Iterative Method (BIM), follow the local gradient of a surrogate model to find a perturbation. While this effectively fools the surrogate, the resulting path is often highly specific to that model's architecture. Tianjin Huang and colleagues in Direction-Aggregated Attack for Transferable Adversarial Examples argue that these gradients are "noisy" and overfit to the local geometry.
To solve this, the research direction has shifted toward regularizing the optimization process. Instead of trusting a single gradient point, these methods look at the "neighborhood" of the image or the "history" of the optimization to find directions that are more likely to be universal across models.
Stabilizing Gradients via Momentum and Variance
One of the most effective ways to prevent overfitting to a surrogate model is to smooth the gradient update. Chuanbiao Song and Jiadong Lin demonstrated in Nesterov Accelerated Gradient and Scale Invariance for Improving Transferability of Adversarial Examples that incorporating Nesterov Accelerated Gradient (NAG) allows the attack to "look ahead" during the optimization. This prevents the attack from getting stuck in poor local maxima that are unique to the surrogate model.
Building on this, Xiaosen Wang and Kun He introduced variance tuning in Enhancing the Transferability of Adversarial Attacks Through Variance Tuning. By calculating the gradient variance in the neighborhood of the current iterate, they can adjust the update step to favor directions with lower variance—essentially seeking out "flatter" regions of the loss landscape that are more likely to transfer to other models.
Input Invariance and Learned Transformations
If an adversarial perturbation is only effective for a specific image at a specific scale, it is unlikely to transfer. A major theme in this research is forcing the attack to be invariant to various transformations.
- Translation Invariance: Tianyu Pang and Yinpeng Dong proposed the Translation-Invariant Attack, which optimizes the perturbation over a set of shifted images. This makes the attack robust against defenses that use image shifting or cropping.
- Scale Invariance: Chuanbiao Song utilized scale-invariant methods to ensure the perturbation remains harmful even if the target model resizes the input.
- Learned Transformations: Moving beyond manual shifts, Weibin Wu and Michael Lyu developed ATTA (Improving the Transferability of Adversarial Samples With Adversarial Transformations). ATTA uses a neural network to learn which transformations are most likely to "break" an adversarial attack and then trains the adversarial example to be robust against those specific transformations.
Adaptive Ensembles and Cross-Architecture Transfer
When multiple surrogate models are available, attackers often use an ensemble. However, not all models in an ensemble are equally helpful. Bin Chen and colleagues addressed this in An Adaptive Model Ensemble Adversarial Attack for Boosting Adversarial Transferability. Their method, AdaEA, dynamically weights surrogate models based on their transferability and filters out "conflicting" gradients where different models disagree. This is particularly effective for transferring attacks between fundamentally different architectures, such as from Convolutional Neural Networks (CNNs) to Vision Transformers (ViTs).
The Power of Modularity
The defining characteristic of this research direction is orthogonality. These techniques—momentum-based optimizers, input transformations, and ensemble strategies—address different facets of the overfitting problem.
As noted by Tianjin Huang, these methods are mutually compatible. An attacker can combine Nesterov momentum with translation invariance and variance tuning to create a "compound" attack. This modularity has led to a cumulative increase in attack potency, revealing that even state-of-the-art "robust" models are highly vulnerable to adversarial examples that have been properly regularized during their creation.