Adversarial debiasing (Zhang et al.)
Algorithms AIMech:ZhangAdversarialDebiasing DRAFT
Aliases: Zhang projected-gradient adversarial debiasing
An adversarial learning algorithm that trains a task predictor alongside a protected-attribute adversary, modifying the predictor gradient with a projection term and an opposing adversary-loss term to target a selected statistical fairness condition.
Identity Notes & Concept Scope
This record isolates Zhang, Lemoine and Mitchell’s projection-based method. It is not the whole family of adversarial fairness approaches, a generative adversarial network, or generic gradient reversal. Different adversary inputs target different fairness constraints rather than defining synonyms for those criteria. All 442 pinned AIO labels/scoped synonyms and 104 preceding native identities were compared, together with hidden/ignored-inclusive searches of both checkouts; source citations and pending frontier mentions were distinguished from existing concept identities.
Claims & Grounded Evidence
2 Claims
Mechanism
#method-scope
Train the adversary to predict the protected attribute from differentiable predictor outputs, optionally also using true labels for equalized odds or a label-restricted training subset for equal opportunity. For the predictor, remove the component of its task-loss gradient along the adversary-loss gradient and subtract a weighted adversary-loss gradient from the update gradient. The source defines a zero projection when the adversary gradient is zero.
Limitation
#projection-stability-limit
The FR-Train authors explicitly omit the Zhang et al. projection term, reporting that it worsened training stability in their experiments. This is a setting-dependent design observation, not evidence that projection-based adversarial debiasing always fails.
⚠️ Curated Limitations & Negative Results
- The source’s convergence argument assumes convexity/concavity, differentiability and convergence; it explicitly notes that the convexity/concavity assumptions usually fail for neural networks.
- Turning adversarial indistinguishability into a population fairness condition requires a sufficiently expressive and well-optimized adversary. Failure of a restricted adversary to detect dependence is not proof of general independence.
- Training can be unstable and is sensitive to optimization settings and loss weighting; the method does not guarantee convergence or exact finite-sample fairness.
- The projection describes an ideal gradient construction. This record does not certify every finite-step or adaptive-optimizer implementation, and omits numerical benchmark claims.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:SupervisedLearning
Supervised Learning
|
related_match | Modifies gradient-based supervised predictive training for classification or regression. The algorithm relates to the learning paradigm and is not mapped as a generative model merely because it uses an adversary. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Brian Hu Zhang, Blake Lemoine, Margaret Mitchell
|
primary paper
full_text
|
ID: arXiv:1801.07593
sha256:0b59ae87abf1…
|
stability-followup
|
Yuji Roh, Kangwook Lee, Steven Whang, Changho Suh
|
primary paper
full_text
|
ID: https://proceedings.mlr.press/v119/roh20a.html
sha256:eeb654425e3d…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T05:03:15Z
Updated: 2026-10-11T05:03:15Z
Source File:
data/curated/ZhangAdversarialDebiasing.yaml