Unbiased Model Prediction Without Using Protected Attribute Information

Puspita Majumdar; Surbhi Mittal; Saheb Chhabra; Mayank Vatsa; Richa Singh

cs.CVMar 31, 2026

Unbiased Model Prediction Without Using Protected Attribute Information

Authors: Puspita Majumdar, Surbhi Mittal, Saheb Chhabra, Mayank Vatsa, Richa Singh

Organizations: IIIT-Delhi, New Delhi, India, 110020

Abstract

The problem of bias persists in the deep learning community as models continue to provide disparate performance across different demographic subgroups. Therefore, several algorithms have been proposed to improve the fairness of deep models. However, a majority of these algorithms utilize the protected attribute information for bias mitigation, which severely limits their application in real-world scenarios. To address this concern, we have proposed a novel algorithm, termed as Non-Protected Attribute-based Debiasing (NPAD) algorithm for bias mitigation, that does not require the protected attribute information. The proposed NPAD algorithm utilizes the auxiliary information provided by the non-protected attributes to optimize the model for bias mitigation. Further, two different loss functions, Debiasing via Attribute Cluster Loss (DACL) and Filter Redundancy Loss (FRL) have been proposed to optimize the model for fairness goals. Multiple experiments are performed on the LFWA and CelebA datasets for facial attribute prediction, and a significant reduction in bias across different gender and age subgroups is observed.

Figures & tables

Figure 1: Illustration of model training using (a) conventional approaches, (b) conventional debiasing approaches, and (c) proposed Non-Protected Attribute-based Debiasing (NPAD) algorithm. The proposed algorithm does not require the protected attribute information for bias mitigation.

Figure 2: Block diagram illustrating model training using the proposed NPAD algorithm. In the first stage, disparity is computed by evaluating the model trained for predicting attribute $\mathbf{A}_{i}$ across the classes of other non-protected attributes. Next, the non-protected attributes are selected based on the disparity and independence. In the second stage, model is optimized using the proposed DACL and FRL functions for bias mitigation.

Figure 3: A toy example of the feature visualization of a model trained for shape prediction on the testing set. Solid and Dotted represent classes of the non-protected attribute and Color represents the protected attribute. (a) Shows the biased feature representations of the model (trained using conventional approaches). (b) Shows the unbiased feature representations of the model (trained using the proposed NPAD algorithm). (c) Shows the unbiased feature representations of the model across the unknown protected attribute (assuming the non-protected attribute is fully correlated with the protected attribute).

Figure 4: Sample images of the (a) LFWA [ 29 ] and (b) CelebA [ 30 ] datasets. Images of both datasets are collected in unconstrained environmental settings with variation in pose, illumination, and the degree of occlusion.

	Algorithm		$P_{M}$ ( Predicted )		$P_{NM}$ ( Predicted )
Ground Truth			B	NB	B	NB
	BMT	B	0	766	0	363
	BMT	NB	0	4654	0	1097
	NPAD (One)	B	553	213	292	71
	NPAD (One)	NB	253	4401	62	1035
	NPAD (Two)	B	511	255	289	74

Table 1: Confusion matrix for ‘Bangs’ prediction (Bangs as ‘B’ and Not Bangs as ‘NB’) across gender subgroups on the LFWA dataset.

Attribute	Algorithm	Across Gender					Across Age
		Accuracy $\uparrow$			DoB $\downarrow$	OPE $\downarrow$	Accuracy $\uparrow$			DoB $\downarrow$	OPE $\downarrow$
		$P_{M}$	$P_{NM}$	Overall	DoB $\downarrow$	OPE $\downarrow$	$P_{Y}$	$P_{NY}$	Overall	DoB $\downarrow$	OPE $\downarrow$
Big Nose	BMT	77.93	75.21	77.35	1.36	2.73	68.78	79.55	77.35	5.38	10.77
	PAD	83.21	82.47	83.05	0.37	0.74	80.26	84.15	83.36	1.94	3.90
	LfF	81.01	77.94	80.36	1.53	3.07	75.19	81.69	80.36	3.24	6.49
	NPAD (One)	82.99	81.92	82.76	0.53	1.07	78.62	83.82	82.76	2.60	5.21

Table 2: Performance of the proposed NPAD algorithm (%) on three attributes of the LFWA dataset. The result closest to PAD is highlighted in bold.

Figure 5: t-SNE visualization using BMT (top row) and the proposed NPAD algorithm (bottom row) corresponding to attributes ‘Big Nose’ and ‘Black Hair’ of the LFWA dataset.

Attribute	Algorithm	Across Gender					Across Age
		Accuracy $\uparrow$			DoB $\downarrow$	OPE $\downarrow$	Accuracy $\uparrow$			DoB $\downarrow$	OPE $\downarrow$
		$P_{M}$	$P_{NM}$	Overall	DoB $\downarrow$	OPE $\downarrow$	$P_{Y}$	$P_{NY}$	Overall	DoB $\downarrow$	OPE $\downarrow$
Wearing Necktie	BMT	81.97	99.93	92.99	8.98	17.96	96.86	80.94	92.99	7.96	15.92
	PAD	91.29	99.90	96.57	4.30	8.61	98.25	92.53	96.86	2.86	5.71
	LfF	81.97	99.93	92.99	8.98	17.96	96.86	80.94	92.99	7.96	15.92
	NPAD (One)	90.71	99.89	96.34	4.59	9.19	97.93	91.40	96.34	3.26	6.53

Table 3: Performance of the proposed NPAD algorithm (%) on three attributes of the CelebA dataset. The result closest to PAD is highlighted in bold.

Attribute	First Non-protected Attribute	Second Non-protected Attribute
Wearing Necktie	0.19	0.42
Sideburns	0.79	0.02
Gray Hair	0.11	0.07

Table 4: Correlation between the non-protected attributes used for debiasing with the protected attribute ‘Gender.’

Figure 6: Mean overall classification accuracy corresponding to all the attributes (excluding male and young ) using BMT, PAD, LfF, and NPAD algorithms on the LFWA and CelebA datasets. PAD (Gender) and PAD (Age) indicates the performance when optimization is done using gender and age information, respectively.

Figure 7: Mean DoB and mean OPE corresponding to all the attributes (excluding male and young ) using BMT, PAD, LfF, and NPAD algorithm on the (a) LFWA and (b) CelebA datasets.

	BMT	NPAD (One)	NPAD (Two)
LFWA	8.64	6.40	6.25
CelebA	8.39	7.12	7.03

Table 5: Mean OPE (%) for all the attributes (excluding male and young) on the LFWA and CelebA datasets across the intersectional subgroups of gender and age .

Figure 8: Cross dataset experiment: mean OPE (%) corresponding to all the attributes (excluding male and young) using BMT and NPAD on (a) LFWA and (b) CelebA datasets.

Figure 9: Mean DoB and mean OPE of the case studies corresponding to the three attributes of the CelebA dataset across gender subgroups. Comparison is performed with mean NPAD.

Figure 10: Overall accuracy of the case studies corresponding to the three attributes of the LFWA dataset. Comparison is performed with mean NPAD.

Loss	Accuracy $\uparrow$			DoB $\downarrow$	OPE $\downarrow$
Loss	$P_{M}$	$P_{NM}$	Overall	DoB $\downarrow$	OPE $\downarrow$
DACL	82.86	81.23	82.51	0.81	1.63
FRL	78.93	76.10	78.33	1.41	2.83
Proposed	82.99	81.92	82.76	0.53	1.07

Table 6: Performance of the model (%) by ablating individual loss terms of the NPAD algorithm (one attribute) for ‘Big Nose’ prediction on the LFWA dataset across different gender subgroups.

Figure 11: DoB and OPE of the model when the independence between non-protected attributes is not considered. The performance is shown for ‘Black Hair’ prediction across age subgroups on the LFWA dataset.

Algo	DoB $\downarrow$		OPE $\downarrow$		Overall Accuracy $\uparrow$
Algo	Gender	Age	Gender	Age	Overall Accuracy $\uparrow$
ARL	4.70	3.34	9.41	6.68	80.33
Proposed	3.67	2.51	7.35	5.03	84.50

Table 7: Comparison with existing algorithm on the LFWA dataset. Average over all attributes (excluding protected attributes - male and young) is reported.

Explore similar work

Dual-Branch Cross-Projection Debiasing through Diffusion-based Disentanglement
Jun 23, 2026Xiangqian Zhao, Xinyang Jiang, Zhipeng Xu +5Disentangled Representation Learning Debiased ML
Xidian University · Shaanxi, China · Microsoft Research Asia +1
Debiasing Without Protected Attributes: Latent Concept Erasure from Textual Profiles
Jun 10, 2026Shun Shao, Zheng Zhao, Anna Korhonen +2Debiased ML Algorithmic Fairness
University of Cambridge · University of Edinburgh · University of Groningen +1
CAFP: A Post-Processing Framework for Group Fairness via Counterfactual Model Averaging
Apr 8, 2026Irina Arévalo, Marcos OlivaAlgorithmic Fairness Group Fairness
School of Civil Engineering, Universidad Politecnica de Madrid, Profesor Aranguren 3, Madrid, 28040, Spain. · Bain&Company, Velazquez, Madrid, 28006, Spain.

Jun 23, 2026J/K move · Enter open · S save

Dual-Branch Cross-Projection Debiasing through Diffusion-based Disentanglement

Xiangqian Zhao, Xinyang Jiang, Zhipeng Xu, Lingfeng He+4

Xidian University · Shaanxi, China · Microsoft Research Asia+1

Disentangled Representation Learning Debiased ML Algorithmic Fairness Spurious Correlation Robustness

Open paper AskPDF arXiv

Foundation models trained on biased datasets often rely on spurious correlations between target labels and non-causal attributes, resulting in poor generalization on minority groups. Bias mitigation remains challenging due to two fundamental issues. First, when group labels are unavailable, existing group-unsupervised methods typically infer spurious attributes implicitly from model behavior, making it difficult to identify spurious factors that are semantically aligned with real-world biases. Second, even with pseudo spurious supervision, most existing debiasing methods follow a single-branch design that operates within a single shared feature space, where target and spurious attributes are intrinsically entangled. To address the first challenge, we introduce Confidence-guided Bias Concept Mining (CBCM), which leverages diffusion-disentangled, semantically grounded concept representations to identify reliable spurious attributes without attribute annotations. To address the second challenge, we propose Dual-branch Cross-projection Debiasing (DCD), a prompt-tuning framework that separates target and spurious representations into two branches and explicitly removes spurious information through cross null-space projection while preserving target-relevant semantics. Extensive experiments on four benchmark datasets show that our method achieves state-of-the-art worst group accuracy among group-unsupervised approaches, while tuning at most 0.22% of the model parameters. The source code is available in the supplementary materials.

Abstract

Explore similar work

Dual-Branch Cross-Projection Debiasing through Diffusion-based Disentanglement

Debiasing Without Protected Attributes: Latent Concept Erasure from Textual Profiles

CAFP: A Post-Processing Framework for Group Fairness via Counterfactual Model Averaging