Many organizations release model features or run prediction services for a population whose sensitive traits, such as sex, race or age, were never collected. They often treat that absence as protection. It is weak protection. An attacker who holds a different, public dataset that does include those traits can train a predictor for them and then use domain adaptation to carry it over to the unlabeled population. Domain adaptation means reshaping a model's internal features so that a predictor trained on one data distribution still works on another. The fairness literature already does exactly this for a benign reason: FAIRDA and its companion Joint Adversarial Learning for Cross-domain Fair Classification align a labeled source domain with an unlabeled target domain so that they can infer the target's missing sex or age labels, and they then debias using those inferred labels. The same step, run by someone else, is an attribute inference attack.
This proposal makes that threat measurable and builds a defense against it. The defense is a feature-release encoder trained against attackers who themselves run domain adaptation, rather than against a fixed classifier. Existing privacy filters, such as adversarially learned obfuscation, Noisy ARL and DRIP, are trained and evaluated inside one labeled domain against an attacker that does not adapt. We expect such filters to look private under their own evaluation and to leak once the attacker adapts to the released target features. The idea in two sentences: treat "attacker plus domain adaptation" as the adversary you train against, and measure privacy as the accuracy of the best adapted attacker. The defender's leverage comes from denying the attacker's alignment step its payoff, which means shaping the released features so that sensitive-attribute structure learned on the source does not carry over to the target, while utility structure still does.
Core Claim
We make two claims. The first is empirical. On standard cross-domain splits, an attacker who applies adversarial domain adaptation to a source-trained attribute predictor recovers substantially more sensitive information from released target features than a source-only attacker does. This holds even when the features were produced by a privacy filter that was trained with source labels. The second claim is methodological. Training the release encoder against a panel of adapting attackers closes most of that gap at a modest cost in utility. The panel draws on adversarial alignment, kernel discrepancy matching and self-training, and it is scored on a small consented target audit set.
No paper in the file measures cross-domain attribute inference as an attack. FAIRDA reports only fairness gains: 21.2% better demographic parity on CelebA at a 0.1% accuracy cost. That result shows the transfer step works well enough to be useful, but the file contains no reported attack-success numbers for it. We found no reported result on Lacuna for cross-domain attribute inference accuracy.
Threat Model and Mechanism
The defender holds raw target data $X_T$ and utility labels $U_T$. It releases $Z_T = O(X_T)$, either as features or through a split-computation service in the style of Noisy ARL, and it can query or reuse the encoder $O$. The attacker holds a public source set $(X_S, S_S)$ with sensitive labels, can compute $Z_S = O(X_S)$, and observes $Z_T$. From these the attacker runs some adaptation procedure $a$ to obtain a predictor $h_a$ of $S$ on the target. GAN-Based Domain Inference Attack supports the assumption that attackers can find a suitable source: it shows black-box models leak which domain they were trained on, and that choosing an auxiliary dataset this way improves downstream model-inversion accuracy by up to 30%.
The defender does not have target sensitive labels. This is what separates our setting from in-domain obfuscation, and it constrains the design. Following the partial-label setting of Towards Group Robustness in the presence of Partial Group Labels, we assume a small consented audit subset $A_T$ of target users who disclose $S$. We sweep the size of this subset down to zero. At zero, the defender falls back on its own adapted pseudo-labels. The training objective is
$$\min_{O,T}\; \mathcal{L}_{\text{util}}\big(T(O(X_T)), U_T\big) \;-\; \lambda \,\min_{a \in \mathcal{P}} \mathcal{L}_{\text{priv}}\big(h_a(O(X_{A_T})), S_{A_T}\big),$$
The terms are defined as follows. $O$ is the release encoder and $T$ the utility head. $\mathcal{P}$ is the panel of adaptation attackers. $h_a$ is the attribute predictor that attacker $a$ produces from the current source and target features. $\mathcal{L}_{\text{priv}}$ is that predictor's cross-entropy on the audit set. $\lambda$ trades utility for privacy. Taking the minimum over the panel means the encoder is penalized by whichever adapted attacker currently succeeds best.
We do not differentiate through each attacker's full adaptation run. Instead, attackers are updated for a few alternating steps between encoder updates, and they are periodically re-initialized so the encoder cannot overfit to one attacker's current state. DAMIA reports that adversarial adaptation methods failed to converge when used for privacy, while discrepancy-based and reconstruction-based methods succeeded. We therefore inject Gaussian noise into $Z$, as Noisy ARL does to stabilize its minimax game. We report the convergence rate across seeds as a first-class result.
The mechanism we expect the encoder to discover draws on a known weakness of domain adaptation. When class-conditional structure differs between domains, global alignment matches the wrong classes together and hurts target accuracy, a failure known as negative transfer. A good defense should keep the utility classes aligned across domains while making the sensitive-attribute clusters of the target land in "wrong" places relative to the source. This induces negative transfer only for $S$. We will test whether this is what happens by measuring, separately for $U$ and for $S$, how closely source and target cluster centroids line up in the released features.
Research Program
The work has three stages. First, we build a reusable leakage audit. For any released encoder, it reports sensitive-attribute accuracy for a source-only attacker, a gradient-reversal adversarial adaptation attacker of the kind described by Ren et al., a kernel-discrepancy (MMD) attacker, a self-training attacker, and an oracle trained on target labels as a ceiling. The gap between the source-only attacker and the best adapted attacker is the "adaptation leakage," the quantity no existing paper reports.
Second, we run the audit on published defenses, retrained on the source data. This checks whether the vulnerability exists. If adaptation leakage is near zero, the threat is overstated. The project then becomes a negative result showing that source-trained filters already transfer, and we report that honestly.
Third, we train the panel defense, ablate its ingredients, and test it against attackers withheld from the panel.
Baselines
The naive reference point is releasing unprotected features from a standard utility model. It shows how much sensitive information the population leaks when nobody intervenes. Every attacker in the audit is run against it.
The first strong competitor is a source-trained in-domain obfuscator. We retrain Noisy ARL using source sensitive labels and apply it unchanged to the target. It is the strongest published in-domain method in the file: on FairFace it reports a privacy-utility gap $\Delta$ of 72.74%, where $\Delta$ is utility accuracy minus privacy-attacker accuracy, against 65.96% for MaxEnt. DRIP reports 89% age accuracy with gender inference reduced to 61% on UTKFace, and it serves as a second in-domain comparator. These numbers were measured against non-adapting attackers within one domain. We will reproduce them in-domain to validate our implementation, then re-evaluate them cross-domain.
The second strong competitor is pseudo-label scrubbing, which is FAIRDA's recipe turned to privacy. The defender adapts a source attribute predictor to the target, infers target labels, and trains an adversarial obfuscator against those inferred labels. This is exactly a one-round version of our method with a single attacker, and it is the most dangerous competitor. We also include DAMIA-style discrepancy adaptation as a defense, using its successful discrepancy variant, to see whether adaptation-based defenses against membership inference carry over to attribute inference.
For fairness, all methods share the same encoder backbone, the same split point, and the same data. They receive an equal hyperparameter search budget, selected on a utility validation set plus the audit set, never on the evaluation set. The ablations are a panel of size one using only adversarial adaptation, a panel without noise injection, and an audit-set size of zero.
Beating the baselines means lowering the best held-out adapted attacker's accuracy by at least a few points while matching utility within one point. Equivalently, it means a higher $\Delta$ computed with the strongest adapted attacker, consistently across seeds and splits.
Why not just…?
Why not just apply differential privacy noise to the released features? Calibrated noise damages every direction in feature space equally, including the ones the utility task needs. Cleaning the Null Space reports that on the scene benchmark, at matched utility error, the Laplace mechanism achieved about 31% "complete privacy" against about 81% for a targeted projection. We will still include a feature-noise curve matched on utility, since noise also blunts adaptation and could be competitive on low-dimensional tabular data such as Adult.
Why not just train an obfuscator on the labeled source and ship it? This is the source-trained Noisy ARL baseline, and our central hypothesis is that it fails. Its attacker was never allowed to adapt, and nothing forces its sensitive-attribute suppression to hold on target features, which the encoder was not trained to protect. The audit settles this directly. If the adapted attacker's gain over a source-only attacker is small against this baseline, our method is unnecessary and we will say so.
Why not just pseudo-label the target and scrub, as FAIRDA does? This is a real threat to the proposal. Our reason to expect a residual gap is that scrubbing against one predictor removes the directions that predictor uses. A differently adapted attacker, for example a self-training or kernel-matching attacker, may find other directions, much as FairVGNN found that masking features by their initial correlation fails once correlations are regenerated downstream. The held-out-attacker experiment decides this question. If one-round pseudo-label scrubbing matches the panel defense on attackers held out from both methods, we will report the simpler method as the answer.
Why not use discrepancy-based adaptation to avoid adversarial instability? DAMIA found that this works for membership inference. We therefore include kernel discrepancy matching both as an attacker and as a defense variant, and we report convergence rates for every adversarial component, rather than assuming adversarial training is required.
Proposed Experiments
We use cross-domain splits already established in the file. For CelebA split by age, as in FAIRDA, the utility task is smiling or attractiveness and the sensitive attribute is sex. For faces across datasets, FairFace serves as the labeled source and UTKFace as the target, with UTKFace's gender and ethnicity labels hidden except in the audit and evaluation sets. For tabular data, we use Adult split by country or working class and COMPAS split by age, following FAIRDA. For each split we report utility accuracy, sensitive-attribute accuracy and balanced accuracy for every attacker, and adaptation leakage. We also report $\Delta$ computed with the strongest attacker, and an intersectional variant for UTKFace ethnicity by gender, since the file notes that nearly all prior formulations treat only binary traits.
Three checks are decisive. The first is whether adaptation leakage is materially above zero for the unprotected release and for the source-trained filters; this establishes whether the threat is real. The second is whether the panel defense suppresses a held-out adapted attacker better than pseudo-label scrubbing at matched utility; this establishes whether the method is needed. The third is how performance changes as the audit set shrinks to zero, which tells practitioners how much consented disclosure the defense requires.
A secondary analysis tests the negative-transfer mechanism by comparing how well source and target clusters for $U$ and for $S$ line up in the released features. Compute stays within a single-GPU budget, since the backbones follow Noisy ARL's ResNet-18 split configuration.
References
- Fair classification via domain adaptation: A dual adversarial learning approach
- Adversarially Learned Representations for Information Obfuscation and Inference
- Noisy adversarial representation learning for effective and efficient image obfuscation
- Data-Driven Regularized Inference Privacy
- GAN-Based Domain Inference Attack
- Invariant Learning with Partial Group Labels
- Leveraging Domain Adaptation as a Defense Against Membership Inference Attacks
- Adversarial Domain Adaptation for Classification of Prostate Histopathology Whole-Slide Images
- Cleaning the Null Space: A Privacy Mechanism for Predictors
- Improving Fairness in Graph Neural Networks via Mitigating Sensitive Attribute Leakage