Hard-Label Black-Box Adversarial Attacks and Model Extraction

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

Version 1 of 2222 papers244 conceptsfrozen 2026-10-04

In the world of machine learning security, a "black-box" model is one where the internal weights and architecture are hidden from the user. However, many real-world APIs go a step further: they provide only a hard label (e.g., "This is a cat") rather than a full probability distribution (e.g., "90% cat, 10% dog").

This creates a massive information bottleneck. Without probability scores (soft labels), an attacker cannot see how close they are to the decision boundary or calculate gradients to guide their search. Researchers have developed a suite of techniques to bypass this restriction, effectively "feeling out" the shape of a model's logic through clever probing and geometric heuristics.

The Optimization Challenge of Discrete Feedback

Traditional adversarial attacks rely on gradients—mathematical indicators of which direction to move an input to change the output. In a hard-label setting, the loss surface is flat and discontinuous; moving an image pixel slightly doesn't change the output label, so the gradient appears to be zero everywhere.

To solve this, Minhao Cheng, Cho-Jui Hsieh, and their colleagues introduced Sign-OPT. Instead of trying to calculate the exact gradient magnitude, they estimate only the sign of the gradient. By querying the model in multiple directions around a point, they can determine which direction is most likely to cross the decision boundary. This transforms a discrete problem into a continuous optimization of the distance to the boundary.

Other researchers, such as Bo Li and Jiawei Zhang, refined this with the Progressive-Scale Boundary Attack (PSBA). They discovered that there is an "optimal scale" for these probes—too small and you see no change, too large and you lose precision. Their method adaptively scales the search to find adversarial examples with significantly fewer queries.

Geometric Heuristics and the Bridge Approach

Intuition suggests that the decision boundary of a neural network is a complex, high-dimensional surface. However, recent work like the Two-third Bridge Approach (TtBA) by Feiyang Wang exploits a specific geometric property. By identifying a "bridge" point—a weight that marks the transition into an adversarial region—they found that a simple $2/3$ scaling factor consistently points toward the most efficient perturbation path. This reduces the need for expensive binary searches along every possible direction.

Extracting the Model Logic

Beyond just fooling a model, adversaries often want to "steal" or clone it. In a hard-label environment, this is difficult because the attacker doesn't know which classes the model almost predicted.

Yixu Wang and Hong Liu addressed this with the Black-Box Dissector. Their insight was to strategically erase parts of an input image. By removing the most discriminative features of the primary class (e.g., erasing a cat's ears), they force the victim model to reveal its "second-best" guess. This "re-mining" of hidden class information allows the attacker to reconstruct the victim's internal similarity structures even when the API only provides a single top-1 label.

Similarly, Kejiang Chen demonstrated in DFHL-RS that it is possible to steal not just a model's accuracy, but also its adversarial robustness. By generating "High-Entropy Examples"—inputs specifically designed to sit on the edge of uncertainty—attackers can map out the defensive boundaries of a model without ever seeing the original training data.

Query Efficiency as the Ultimate Metric

Because commercial APIs often have rate limits or costs per query, the success of these attacks is measured by query efficiency.

  • ActiveThief by Soham Pal uses active learning to select only the most informative samples to query, minimizing the "budget" needed to clone a model.
  • BOSH uses Bayesian Optimization to explore multiple solution paths simultaneously, preventing the search from getting stuck in the local minima common in hard-label landscapes.

By treating the hard-label model as a "distance oracle" rather than a differentiable function, this research direction has shown that the information bottleneck of discrete labels is far more porous than previously assumed. Even the most restrictive interfaces can leak enough geometric information to allow for high-fidelity model extraction and precise adversarial manipulation.

Papers in this version

Showing 60 of 222 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 160 papers
  2. Version 1 · 2026-10-04 · 222 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_hard_label_black_box_constraint_5409e8f7,
  title = {hard-label black-box constraint},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hard-label-black-box-constraint/rdr_5409e8f789d041dab128b387d5b6ec57},
  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_hard_label_black_box_adversarial_attacks_5409e8f7_v1,
  title = {Hard-Label Black-Box Adversarial Attacks and Model Extraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hard-label-black-box-adversarial-attacks-and-model-extraction/rdr_5409e8f789d041dab128b387d5b6ec57/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}