Adversarial Transferability through Loss Landscape Flatness

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 265 papers137 conceptsfrozen 2026-10-04

In the cat-and-mouse game of AI security, an attacker often lacks access to the internal weights of their target. To bypass this, they craft adversarial examples on a local "surrogate" model, hoping the perturbation will "transfer" to the hidden victim. However, most attacks are brittle: a perturbation that perfectly fools a ResNet might be ignored by a Vision Transformer.

Recent research suggests that this failure is a geometric problem. Standard attacks often converge to "sharp" peaks in the loss landscape—narrow spikes of high error that are unique to the surrogate model’s specific weights. To create truly universal attacks, researchers are moving away from finding the highest loss and instead seeking the flattest loss.

The Geometry of Generalization

The core intuition behind this research direction is an analogy to model training. Just as flat minima in a loss landscape help a model generalize from training data to unseen test data, flat regions of high loss help an adversarial attack generalize from a surrogate model to an unseen target.

When an attack targets a sharp peak, even a slight shift in the decision boundary—caused by a different architecture or even simple quantization—can move the input out of the "danger zone." As Yanbo Fan and colleagues argue in Boosting the Transferability of Adversarial Attacks with Reverse Adversarial Perturbation, seeking adversarial examples within flat regions makes them less sensitive to these structural differences.

Finding the Plateaus

If the goal is to find broad plateaus of high loss rather than narrow needles, how do we guide the optimization? Several distinct strategies have emerged:

  • Bayesian Ensembles: Instead of attacking one set of weights, Qizhang Li and Steven Y. Guo propose making the substitute model "Bayesian." By treating weights as a probability distribution, the attack effectively targets an infinite ensemble of models. This naturally favors flat regions because a perturbation must remain effective across a wide range of weight variations, as detailed in Making Substitute Models More Bayesian Can Enhance Transferability of Adversarial Examples.
  • Adversarial Weight Tuning (AWT): Rather than keeping the surrogate model static, Jiahao Chen and Feng Zhou suggest adaptively tuning the surrogate’s weights during the attack. In Enhancing Adversarial Transferability with Adversarial Weight Tuning, they demonstrate that optimizing for "flat local maxima" in both the input and parameter spaces simultaneously creates perturbations that are far more robust to architectural shifts.
  • Smoothing the Landscape: Some domains, like skeletal action recognition, have notoriously "jagged" landscapes. Yunfeng Diao shows in TASAR: Transferable Attack on Skeletal Action Recognition that smoothing these surfaces—essentially blurring the loss landscape—allows the optimizer to ignore model-specific spikes and find the underlying vulnerabilities shared across different architectures.

The Limits of Flatness

While the "flatter is better" heuristic is powerful, it is not a complete theory. Mingyuan Fan and Xiaodan Li have begun to formalize this via Transferability Bound Theory. They prove that transferability depends on three distinct factors: the discrepancy between models, first-order flatness (the gradient), and second-order curvature (the Hessian). An attack might reside in a region that looks flat but has high second-order curvature, meaning the gradient changes rapidly nearby, causing the attack to fail.

Furthermore, Linara Adilova and Michael Kamp identify an "Uncanny Valley" in The Uncanny Valley: Exploring Adversarial Robustness from a Flatness Perspective. They found that adversarial examples often settle into regions that are too flat—vast plateaus where the model is confidently and stably wrong. This suggests that while flatness helps an attack transfer, it also makes those attacks harder to detect or mitigate using gradient-based defenses.

Why It Works

The shift toward flatness-aware attacks represents a transition from "overfitting" a single model to "generalizing" across a model distribution. By targeting the geometric properties of the loss surface—specifically low curvature and stable gradients—researchers like Shuyang Hao are proving that even complex, multi-modal systems like Vision-Language Models are vulnerable to these "universal" geometric features. The reader should take away that the most dangerous adversarial examples are not those that maximize error, but those that find regions where error is inevitable and stable across the entire landscape of possible models.

Papers in this version

Showing 60 of 65 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 35 papers
  2. Version 1 · 2026-10-04 · 65 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_loss_landscape_flatness_653ad99a,
  title = {loss landscape flatness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/loss-landscape-flatness/rdr_653ad99a3f07485d810ac4542651f38a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_adversarial_transferability_through_loss_653ad99a_v1,
  title = {Adversarial Transferability through Loss Landscape Flatness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adversarial-transferability-through-loss-landscape-flatness/rdr_653ad99a3f07485d810ac4542651f38a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}