The Equivalence of Adversarial Perturbations and Optimal Transport

18 concepts8 papers1 proposal

Adversarial attacks are typically viewed as localized, malicious glitches—small, norm-bounded perturbations engineered to trick a specific neural network into misclassifying an input. However, a deeper geometric perspective reveals that adversarial manipulation is not merely a collection of isolated hacks. Instead, it behaves like a macroscopic physical process: the systematic transportation of probability mass from one data distribution to another.

By framing adversarial machine learning through the lens of Optimal Transport (OT), researchers have uncovered a profound mathematical equivalence between fooling a model and solving a transport plan. This connection provides a rigorous framework for calculating the absolute limits of adversarial robustness and yields surprisingly efficient algorithms for domain adaptation.

The Adversary as a Mass Transporter

To understand why optimal transport fits the adversarial setting, we must shift our focus from individual data points to entire class distributions. Imagine two distinct probability distributions representing two classes, such as images of cats and images of dogs. Under normal conditions, a classifier draws a decision boundary between these two distributions.

An adversary's goal is to modify the inputs so that the classifier can no longer distinguish between them. From a distributional perspective, the adversary is trying to "compress" the statistical distance between the two classes until they overlap. As shown by Menon et al. and Cranko et al., when an attacker successfully reduces the integral probability metric between class-conditional distributions below a certain threshold, the classes become fundamentally indistinguishable. At this point, any classifier—regardless of its architecture or training regime—is forced into a non-informative state, relying purely on class base rates rather than informative features.

For classifiers constrained by regularity conditions like Lipschitz continuity, this blurring of boundaries is exactly an optimal transport problem. The adversary acts as a transport agent, shifting the mass of one distribution to mirror the other while minimizing the "work" (perturbation budget) required to do so.

Deriving the Fundamental Limits of Robustness

This transport-based framing allows us to calculate universal, classifier-agnostic lower bounds on adversarial risk. If we know the inherent geometric distance between two class distributions, we can determine exactly when a robust classifier is theoretically impossible.

To formalize this, Pydi et al. characterized adversarial risk using a specific binary cost function within the Monge-Kantorovich OT framework. In their formulation, the cost of transporting a point from one class to another is defined as zero if the two points can be perturbed within an adversarial budget $\epsilon$ to become identical, and one otherwise.

   Class A (Data Distribution)               Class B (Data Distribution)
         [ Mass Pile A ]                           [ Mass Pile B ]
                \                                         /
                 \___ Adversary shifts mass with cost ___/
                      (0 if within budget, 1 otherwise)
                                  |
                                  v
                     [ Indistinguishable Overlap ]

By applying Kantorovich duality to this setting, Bhagoji et al. demonstrated that the minimum possible 0-1 loss under attack is equal to the dual of this transportation problem. If the adversary's perturbation budget exceeds the minimum transport cost required to make the expanded supports of the two distributions overlap, no model can achieve perfect accuracy. This builds on foundational insights from El-Yaniv et al., which established that classification difficulty is inherently bounded by the statistical distance between the target distribution and the adversarial distribution.

Ultimately, this means adversarial robustness has a hard mathematical ceiling. If the geometry of your data allows an adversary to cheaply transport mass from Class A to Class B, the vulnerability is a property of the dataset itself, not a flaw in the model's architecture.

Why Adversarial Attacks are True Transport Maps

The connection between optimal transport and adversarial attacks runs deeper than a shared optimization philosophy. The actual mathematical maps generated by gradient-based adversarial attacks are optimal transport maps.

In optimal transport theory, a map is proven to be optimal if and only if it satisfies a geometric property known as $c$-cyclical monotonicity. Intuitively, this means that if you select a set of points and map them to a set of destinations, you cannot reduce the total transport cost by swapping any of the destinations in a closed loop. The paths chosen by an optimal transport plan never cross in a way that wastes energy.

Cyclically Monotone (Optimal)             Non-Monotone (Inefficient)
     X_1 ---------> Y_1                        X_1 ----. .----> Y_1
                                                        X
     X_2 ---------> Y_2                        X_2 ----' '----> Y_2
   (Paths do not cross)                      (Swapping paths reduces cost)

Asadulaev et al. discovered that small dataset perturbations generated by iterative adversarial attacks inherently exhibit this $c$-cyclical monotonicity. This provides a formal mathematical equivalence: the directional updates calculated via backpropagation to fool a model are structurally identical to the vectors calculated by discrete optimal transport solvers to move mass efficiently. Adversarial perturbations are not random direction glitches; they follow strict geometric principles that align distributions along the shortest path possible.

Turning Attacks into Solutions

Recognizing that adversarial attacks are implicit optimal transport maps allows researchers to flip the script. Instead of using these gradients to break models, we can use them to solve distribution alignment problems, such as domain adaptation.

In domain adaptation, you have a model trained on a source domain (e.g., synthetic images) and want it to perform well on a target domain (e.g., real-world photos). This requires finding an optimal transport map to align the target distribution with the source distribution. Traditional discrete OT solvers are computationally expensive and scale poorly to high-dimensional deep learning datasets.

As demonstrated by Asadulaev et al., we can bypass heavy OT machinery by leveraging standard adversarial attack algorithms like the Fast Gradient Sign Method (FGSM). Because these gradient-based attacks naturally generate $c$-cyclically monotone maps, they can be repurposed as fast, lightweight transport operators. By deliberately perturbing target domain samples so that they satisfy the expectations of a source classifier—a technique referred to as "source fiction"—the model's own adversarial gradients act as the transport vectors needed to bridge the domain gap.

Key Foundational Work

Monge beats Bayes: Hardness Results for Adversarial Training (2018) — contributed 4 concept elements to this direction.
Monge blunts Bayes: Hardness Results for Adversarial Training. (2018) — contributed 2 concept elements to this direction.
Optimal Single-Class Classification Strategies (2006) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_adversarial_optimal_transport_equivalenc_d1a1eab5,
  title = {adversarial-optimal transport equivalence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adversarial-optimal-transport-equivalence/rdr_d1a1eab5aa48447fb2a0f9cd21de3930},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_adversarial_optimal_transport_equivalenc_d1a1eab5_v1,
  title = {adversarial-optimal transport equivalence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adversarial-optimal-transport-equivalence/rdr_d1a1eab5aa48447fb2a0f9cd21de3930/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake