BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance

Haojin Deng; Zhiping Lin; Yimin Yang

cs.AIOct 5, 2026

BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance

Authors: Haojin Deng, Zhiping Lin, Yimin Yang

Organizations: Vector Institute, Toronto, Canada

Abstract

Worst-group accuracy (WGA) evaluates a trained predictor but does not characterize how its frozen backbone behaves when a new head is learned. We introduce BiasFlow, a hook-based toolkit for monitoring class-attribute centroid alignment (IBMI), within-class centroid separation (W-IBMI), and feature-projection sensitivity. IBMI is confounded by class-attribute correlation and is not a measure of causal feature reliance. We pair these diagnostics with BiasFlow Regularization (BFR), a supervised, composable class-conditional centroid-alignment penalty. W-IBMI verifies the quantity BFR optimizes; it is scale dependent and does not independently establish attribute removal. Across the reported small-scale benchmarks, adding BFR improves or preserves mean WGA, with gains up to +26.0 pp on UrbanCars. The principal independent stress test freezes CelebA-Std backbones and trains fresh heads on biased data: BFR+GroupDRO improves WGA from 40.7% to 64.1%, while Male probe accuracy decreases from 92.5% to 72.2%. Attribute information remains recoverable, and cross-task results are mixed. A controlled synthetic-watermark ImageNet experiment additionally improves watermark-shift accuracy by +23.0 pp under matched training. These results support evaluating centroid geometry and resistance to biased head retraining alongside WGA, within the tested protocols.

Figures & tables

Fig. 1: t-SNE of CelebA-Std ResNet-18 features, seed 1 (Blond Hair target, Male attribute). Baseline and BFR panels illustrate different within-class subgroup layouts. BFR reduces the targeted centroid discrepancy; two-dimensional mixing does not establish concept erasure or causal feature use.

Panel of Figure 2 — Fig. 2: BiasFlow workflow and evidence roles. (a) Forward hooks expose centroid geometry and unsigned projection sensitivity. Blue circles and orange triangles denote attribute values, dashed ellipses denote target classes, and open markers denote group centroids. (b) BFR-inv penalizes within-class centroid differences, averaged over annotated attributes during backbone training; W-IBMI is scale-dependent, objective-aligned verification. (c) Frozen-backbone tests fit a fresh task head on biased training features and a separate attribute probe on validation features, then evaluate on test features. Geometry is schematic, not measured data. Centroid alignment and projection sensitivity do not establish concept erasure or a causal intervention-selection rule; predictive evidence is limited to the tested protocols.

Dataset	Arch	Opt	lr	wd	mom	bs	Epochs
CelebA (R18)	R18	SGD	1e-4	1e-4	0.9	128	50
CelebA (R50)	R50	SGD	1e-4	1e-4	0.9	128	50
CelebA-MB	R18	SGD	1e-3	1e-4	0.8	64	50
UrbanCars	R50	SGD	1e-3	1e-3	0.1	128	60
Waterbirds	R50	SGD	1e-3	1e-4	0.9	128	300
Method overrides (applied on top of dataset defaults):

TABLE I: Training hyperparameters per dataset. Method-level overrides (below the rule) apply on top of dataset defaults. wd = weight decay, mom = momentum, bs = batch size.

Method	CelebA R18	CelebA R50	CelebA-MB R18	UrbanCars † R50	Waterbirds R50
ERM	$0.404{\pm}0.008$	$0.559{\pm}0.004$	$0.274{\pm}0.010$	$0.217$ / $0.489$	$0.357{\pm}0.029$
LfF	$0.681{\pm}0.055$	$0.881{\pm}0.003$	$0.940{\pm}0.017$	$0.108$ / $0.421$	$0.650{\pm}0.059$
JTT	$0.609{\pm}0.010$	$0.647{\pm}0.016$	$0.013{\pm}0.014^{*}$	$0.158$ / $0.075$	$0.060{\pm}0.012^{*}$
DFR	$0.502{\pm}0.019$	$0.559{\pm}0.004$	$0.468{\pm}0.017$	$0.212$ / $0.474$	$0.331{\pm}0.004$
DebiAN	$0.513{\pm}0.033$	$0.896{\pm}0.008$	$0.951{\pm}0.014$	$0.089$ / $0.312$	$0.597{\pm}0.034$
GroupDRO	$0.466{\pm}0.023$	$0.650{\pm}0.004$	N/A ‡	$0.044{\pm}0.014$ / $0.372{\pm}0.018$	$0.329{\pm}0.003$

TABLE II: IBMI of trained backbones (reported 3-seed means and standard deviations; UrbanCars reports mean background/object values). IBMI is a class-confounded centroid-alignment summary, not actual feature reliance. DFR freezes its own first-stage backbone, which need not equal the separately trained ERM run. ∗ Low-IBMI cases discussed in the text. N/A marks an unavailable configuration; GroupDRO uses ResNet-50 for the reported CelebA-MB configuration.

Fig. 3: Unsigned AblGap on CelebA-Std and UrbanCars, using the original plotted seed values. AblGap is

|\Delta\mathrm{Acc}_{R}-\Delta\mathrm{Acc}_{U}|

, reported in pp. Larger values indicate more asymmetric projection sensitivity across groups. The bars do not show the sign of either accuracy change, and cannot distinguish exploitation from interference.

		CelebA-Std (ResNet-18)			CelebA-MB (ResNet-18)
Method	Paradigm	Published WGA	Reproduced WGA	+BFR WGA	Reproduced WGA	+BFR WGA
ERM	Baseline	37.9	43.9 $\pm$ 1.2	54.6 $\pm$ 2.6	40.3 $\pm$ 1.6	53.9 $\pm$ 2.7
LfF	Loss-reweight	43.3	58.7 $\pm$ 2.1	61.7 $\pm$ 2.3	49.1 $\pm$ 3.3	67.7 $\pm$ 2.3
JTT	Loss-reweight	81.1	81.1 $\pm$ 0.4	82.2 $\pm$ 0.5	62.7 $\pm$ 3.8	69.3 $\pm$ 2.6
DFR †‡	Head-retrain	92.9	80.6 $\pm$ 2.5	81.9 $\pm$ 1.0	73.9 $\pm$ 0.8	75.7 $\pm$ 1.0
DebiAN	Loss-reweight	52.9	64.4 $\pm$ 3.2	65.4 $\pm$ 5.3	46.9 $\pm$ 4.2	58.4 $\pm$ 2.4

TABLE III: WGA (%) on CelebA-Std (ResNet-18) and CelebA-MB (ResNet-18). 3 seeds, mean

\pm

std. Bold : higher WGA between Reproduced and +BFR for each method. Italics : published number for reference. † Supervised method. ‡ Published DFR (92.9) uses ResNet-50 + sklearn solver; our R18 reproduction uses L-BFGS head, explaining the gap.

Fig. 4: Recorded IBMI and W-IBMI trajectories for ERM and BFR+ERM on CelebA-Std. Pooled centroid alignment can increase as the attribute axis becomes dominated by class composition. W-IBMI tracks the centroid penalty being optimized, rather than independent concept erasure.

	CelebA R18		CelebA R50		Waterbirds R50
Method	Base	+BFR	Base	+BFR	Base	+BFR
ERM	$0.090$	$0.001$	$0.018$	$0.002$	$0.027$	$0.001$
LfF	$0.298$	$0.000$	$0.082$	$0.000$	$0.158$	$0.000$
JTT	$0.037$	$0.001$	$0.008$	$0.001$	$0.013$	$0.004^{\ddagger}$
DFR	$0.063$	$0.001$	$0.018$	$0.002$	$0.026$	$0.001$
DebiAN	$0.154$	$0.001$	$0.060$	$0.001$	$0.116$	$0.001$

TABLE IV: Reported W-IBMI before and after BFR-inv (3-seed means). This scale-dependent first-moment statistic verifies an objective-aligned effect within each feature convention; it cannot rank architectures or certify erasure. ∗ BFR+GroupDRO R50 does not reduce W-IBMI; JTT+BFR Waterbirds retains a larger residual than most other configurations.

Fig. 5: t-SNE of CelebA-MB ResNet-18 features (Smiling target; Male and Young annotated attributes). Color encodes target and Male; shape encodes Young. Baseline panels show subgroup separation; BFR panels illustrate more within-class mixing. BFR averages separate class-conditional penalties for the two supplied attributes. The visualization does not establish joint distribution alignment or lack of attribute recoverability.

		WGA (%)
Method	Paradigm	Published	Reproduced	+BFR
ERM	Baseline	47.7	$49.1{\pm}1.6$	$\mathbf{50.6{\pm}0.5}$
LfF	Loss-reweight	77.2	$55.9{\pm}5.9$	$\mathbf{62.6{\pm}8.1}$
JTT	Loss-reweight	81.5	$80.2{\pm}0.2$	$\mathbf{82.1{\pm}0.2}$
DFR †	Head-retrain	80.7	$81.3{\pm}1.8$	$\mathbf{85.4{\pm}2.2}$
DebiAN	Loss-reweight	—	$62.6{\pm}3.3$	$\mathbf{64.1{\pm}3.0}$

TABLE V: WGA (%) on CelebA-Std (ResNet-50, ImageNet pretrained). 3 seeds, mean

\pm

std, WGA-based model selection. Bold : higher WGA between Reproduced and +BFR for each method. Italics : published number for reference. † Supervised method.

		UrbanCars (R50)		Waterbirds (R50)
Method	Paradigm	ID Acc (base / +BFR)	WGA (base / +BFR)	Avg Acc (base / +BFR)	WGA (base / +BFR)
ERM	Baseline	$\mathbf{97.5{\pm}0.3}$ / $96.4{\pm}0.6$	$56.7{\pm}0.5$ / $\mathbf{82.7{\pm}0.5}$	$91.2$ / $\mathbf{94.0}$	$79.0{\pm}0.1$ / $\mathbf{83.2{\pm}1.4}$
LfF	Loss-reweight	$97.2{\pm}1.7$ / $\mathbf{97.4{\pm}0.9}$	$67.7{\pm}3.6$ / $\mathbf{85.6{\pm}2.8}$	$91.9$ / $\mathbf{92.7}$	$79.4{\pm}1.5$ / $\mathbf{86.5{\pm}1.6}$
JTT ∗	Loss-reweight	$91.6{\pm}1.4$ / $\mathbf{91.9{\pm}1.7}$	$56.7{\pm}2.5$ / $\mathbf{60.1{\pm}2.3}$	$88.1$ / $\mathbf{88.3}$	$80.8{\pm}3.8$ / $\mathbf{80.9{\pm}3.8}$
DFR †	Head-retrain	$86.8{\pm}0.6$ / $\mathbf{88.4{\pm}0.1}$	$78.1{\pm}0.2$ / $\mathbf{83.9{\pm}1.6}$	$\mathbf{93.3}$ / $93.0$	$91.3{\pm}0.5$ / $\mathbf{91.4{\pm}0.5}$
DebiAN	Loss-reweight	$\mathbf{98.1{\pm}1.1}$ / $97.5{\pm}0.3$	$63.7{\pm}2.6$ / $\mathbf{79.6{\pm}2.6}$	$91.0$ / $\mathbf{93.8}$	$80.1{\pm}2.4$ / $\mathbf{87.9{\pm}1.3}$

TABLE VI: WGA (%) and ID Accuracy (%) on scene/object datasets (3 seeds, mean

\pm

std). Bold : higher value between base and +BFR for each reported metric within each method. † Supervised method.

Method	CelebA-Std	CelebA-Std	CelebA-MB	UrbanCars
	(R18)	(R50)	(R18)	(R50)
GroupDRO	$86.3{\pm}0.5$	$88.0{\pm}0.7$	$55.7{\pm}0.5$	$74.1{\pm}1.2$
BFR + GroupDRO	$\mathbf{88.0{\pm}1.5}$	$\mathbf{88.5{\pm}0.5}$	$\mathbf{68.3{\pm}4.4}$	$\mathbf{87.5{\pm}0.2}$
$\Delta$ WGA (pp)	$+1.7$	$+0.5$	$+12.5$	$+13.3$

TABLE VII: GroupDRO vs. BFR+GroupDRO (3 seeds, mean

\pm

std). WGA (%). BFR adds gains to the evaluated supervised GroupDRO baseline.

	ERM	BFR	$\Delta$
OOD accuracy (%, $\uparrow$ )
Val (ID)	$76.0{\pm}0.1$	$76.0{\pm}0.1$	$\phantom{+}0.0$
IN-W	$52.6{\pm}0.6$	$\mathbf{75.6{\pm}0.1}$	$+23.0$
IN-R	$36.2{\pm}0.6$	$36.5{\pm}0.2$	$+0.3$
IN-Stylized	$52.1{\pm}1.0$	$53.2{\pm}0.5$	$+1.1$
mixed_same	$87.2{\pm}0.3$	$87.3{\pm}0.1$	$+0.1$

TABLE VIII: Controlled synthetic-watermark training on ImageNet-1K, ResNet-50. Accuracy rows retain 3-seed means and standard deviations; IN-W denotes the reported watermark-shift evaluation after injected-watermark training, not a canonical natural-training comparison. Geometry rows retain point summaries: W-IBMI is objective aligned and scale dependent; AblGap is unsigned. Accuracy changes are in pp.

Method	Paradigm	Biased WGA ( $\uparrow$ )	Male Probe ( $\downarrow$ )
Method	Paradigm	(base / +BFR)	(base / +BFR)
ERM	Baseline	$40.0$ / $\mathbf{44.6}$	$93.2$ / $\mathbf{80.0}$
LfF	Loss-rewt.	$44.8$ / $\mathbf{60.0}$	$90.6$ / $\mathbf{61.4}$
JTT	Loss-rewt.	$36.5$ / $\mathbf{42.2}$	$93.4$ / $\mathbf{86.6}$
DFR	Head-retrain	$42.2$ / $\mathbf{44.4}$	$91.9$ / $\mathbf{67.5}$
DebiAN	Loss-rewt.	$45.4$ / $\mathbf{52.6}$	$92.8$ / $\mathbf{65.2}$

TABLE IX: Frozen-backbone stress tests on CelebA-Std (R18), three-seed means. Biased WGA is test WGA of a fresh SGD task head fitted on biased training features. Male Probe is test accuracy of an SGD linear attribute classifier fitted on validation features. Each pair is base / +BFR; bold marks higher WGA or lower attribute-probe accuracy. Lower probe accuracy indicates reduced recoverability under this protocol, not certified erasure. Per-seed variability is in Appendix G .

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Mode	ERM	LfF	JTT	DFR	DebiAN
Baseline (no BFR)	38.3	50.9	75.7	79.6	66.5
orth	37.8	62.8	76.9	78.3	66.1
spectral	39.5	40.0	0.6	38.3	46.7
inv	61.1	82.8	78.8	81.1	77.2

Appendix

TABLE S1: BFR mode ablation on CelebA-Std R18, seed=1,

\lambda{=}20

. WGA (%). Bold : best per method.

Method	$\lambda{=}5$	$\lambda{=}10$	$\lambda{=}20$	$\lambda{=}30$	Trend
ERM	48.9	55.6	61.1	64.4	$\uparrow$ monotone
LfF	82.2	82.2	82.8	80.0	peaked at 20
JTT	78.8	78.6	76.1	$0.0^{*}$	$\downarrow$ , collapses at 30
DFR	75.6	79.4	81.1	77.2	peaked at 20
DebiAN	66.1	71.7	77.2	46.7	peaked at 20
Selected $\lambda$ : 1 for JTT; 20 for all others.

Appendix

TABLE S2:

\lambda

sensitivity for BFR- inv on CelebA-Std R18, seed=1. WGA (%), accuracy-based model selection. Bold : selected value per method. ∗ Accuracy-selected checkpoint has WGA

=

0; WGA-selected checkpoint achieves

73.9

Method	Initial WGA (base / +BFR)	Final WGA (base / +BFR)
ERM	$43.9\pm 1.2$ / 54.6 $\pm$ 2.6	$40.9\pm 2.7$ / 42.6 $\pm$ 1.1
LfF	$58.7\pm 2.1$ / 61.7 $\pm$ 2.3	46.9 $\pm$ 0.3 / $46.1\pm 2.1$
JTT	$81.1\pm 0.4$ / 82.2 $\pm$ 0.5	$42.8\pm 2.0$ / 44.8 $\pm$ 2.1
DFR	$80.6\pm 2.5$ / 81.9 $\pm$ 1.0	$43.7\pm 3.3$ / 63.5 $\pm$ 12.7
DebiAN	$64.4\pm 3.2$ / 65.4 $\pm$ 5.3	$43.7\pm 3.6$ / 45.7 $\pm$ 2.5
GroupDRO	$86.3\pm 0.5$ / 88.0 $\pm$ 1.5	$42.6\pm 2.7$ / 44.4 $\pm$ 0.5

Appendix

TABLE S3: Continued-training durability on CelebA-Std: WGA (%) after standard ERM continuation on the original Blond classification task and biased training distribution. Each metric is reported as base / +BFR. Bold marks the higher WGA within each pair.

Fig. S1: WGA over 30 epochs of ERM continuation from each method’s converged checkpoint on the original CelebA-Std Blond classification task. BFR-DFR reaches

63.5\%

final WGA, substantially above DFR without BFR, indicating stronger retained WGA after continued training.

Method	Paradigm	Acc ( $\uparrow$ )	Mean Group Acc ( $\uparrow$ )	WGA ( $\uparrow$ )	Group Disparity ( $\downarrow$ )
ERM	Baseline	$81.1\pm 0.5$ / 81.9 $\pm$ 0.8	$80.3\pm 0.5$ / 81.2 $\pm$ 0.8	$74.2\pm 0.5$ / 76.9 $\pm$ 1.3	$10.4\pm 0.4$ / 6.5 $\pm$ 0.5
LfF	Loss-rewt.	77.8 $\pm$ 2.1 / $77.1\pm 4.6$	77.1 $\pm$ 2.1 / $76.5\pm 4.6$	71.0 $\pm$ 1.5 / $70.2\pm 6.3$	10.9 $\pm$ 0.2 / $11.9\pm 2.4$
JTT	Loss-rewt.	80.1 $\pm$ 0.4 / $79.4\pm 0.3$	79.2 $\pm$ 0.4 / $78.3\pm 0.3$	71.3 $\pm$ 0.5 / $68.0\pm 0.4$	13.1 $\pm$ 0.7 / $16.6\pm 0.9$
DFR	Head-retrain	80.1 $\pm$ 0.3 / $78.6\pm 0.6$	79.3 $\pm$ 0.4 / $78.1\pm 0.6$	$72.7\pm 0.6$ / 73.5 $\pm$ 1.1	$10.8\pm 0.2$ / 7.1 $\pm$ 1.6
DebiAN	Loss-rewt.	80.2 $\pm$ 2.1 / $75.7\pm 5.7$	79.4 $\pm$ 2.2 / $75.2\pm 5.8$	72.9 $\pm$ 2.6 / $69.5\pm 9.7$	$11.3\pm 0.8$ / 10.9 $\pm$ 6.4
GroupDRO	Supervised	$80.4\pm 0.4$ / 82.4 $\pm$ 0.3	$79.8\pm 0.4$ / 82.1 $\pm$ 0.4	$74.0\pm 0.7$ / 79.6 $\pm$ 0.3	$9.5\pm 0.3$ / 5.5 $\pm$ 0.4

Appendix

TABLE S4: Cross-task transfer to Smiling prediction with the Blond/Male backbone frozen and a fresh biased downstream head trained on standard Smiling data. Acc is overall accuracy; Mean Group Acc averages the four Smiling/Male groups; Group Disparity is the best-group minus worst-group accuracy. 3 seeds, mean

\pm

std. Each metric is reported as base / +BFR, and bold marks the better value within each pair.

		ERM			LfF			JTT			DFR			DebiAN			GroupDRO
Attribute	$\|r\|$	Base	+BFR	$\Delta$	Base	+BFR	$\Delta$	Base	+BFR	$\Delta$	Base	+BFR	$\Delta$	Base	+BFR	$\Delta$	Base	+BFR	$\Delta$
Narrow_Eyes	.013	85.3	85.3	$-0.0$	85.1	85.1	$+0.0$	85.2	85.1	$-0.1$	85.1	85.1	$+0.0$	85.1	85.1	$-0.0$	85.3	85.1	$-0.2$
Blurry	.026	95.5	95.1	$-0.4$	95.2	94.9	$-0.3$	95.6	95.0	$-0.5$	95.4	94.9	$-0.4$	95.3	95.0	$-0.3$	95.5	94.9	$-0.5$
Straight_Hair	.063	81.2	79.7	$-1.5$	79.9	79.0	$-0.9$	81.5	79.5	$-2.0$	80.3	79.0	$-1.2$	80.1	79.0	$-1.1$	81.3	79.4	$-1.9$
Pale_Skin	.076	96.0	95.8	$-0.2$	95.9	95.8	$-0.1$	96.0	95.8	$-0.2$	95.9	95.8	$-0.1$	96.0	95.8	$-0.2$	96.0	95.8	$-0.2$
Mouth_Open	.100	74.7	75.4	$+0.7$	68.2	62.9	$-5.3$	74.3	69.6	$-4.7$	73.2	65.7	$-7.4$	74.3	61.2	$-13.2$	74.2	73.9	$-0.3$

Appendix

TABLE S5: Attribute correlation spectrum on CelebA-Std with ResNet-18, averaged over 3 seeds. Backbones are trained for Blond Hair classification with Male as the BFR-targeted spurious attribute, then frozen for linear-probe evaluation across 12 CelebA attributes sorted by

|r(\text{attr},\,\text{Male})|

. Each method reports Base (no BFR) / +BFR /

\Delta

(BFR

-

Base). The bottom row reports Spearman

\rho

between

|r|

and probe accuracy; bold marks the more selective Base/+BFR value within each method.

Fig. S2: Attribute-probe accuracy differences relative to ERM vs. absolute label correlation with Male, for BFR-DFR and BFR-GroupDRO. The saved plotting source uses ERM as the reference for both curves; this is not the paired BFR-minus-corresponding-base comparison in Table S5 . Larger losses occur for several Male-correlated attributes, with exceptions and collateral losses at low correlation. The plot does not isolate the within-method effect of BFR or establish a distinct erased subspace.

Method	Biased test WGA	Male test accuracy	Balanced test WGA
ERM	$40.0\pm 1.5$ / $44.6\pm 5.3$	$93.2\pm 0.1$ / $80.0\pm 0.1$	$83.3\pm 2.9$ / $81.3\pm 5.0$
LfF	$44.8\pm 5.2$ / $60.0\pm 3.5$	$90.6\pm 2.5$ / $61.4\pm 7.1$	$80.2\pm 4.5$ / $82.6\pm 2.6$
JTT	$36.5\pm 2.2$ / $42.2\pm 0.6$	$93.4\pm 0.3$ / $86.6\pm 0.4$	$81.3\pm 1.3$ / $79.6\pm 1.8$
DFR	$42.2\pm 1.0$ / $44.4\pm 3.6$	$91.9\pm 0.2$ / $67.5\pm 1.1$	$80.2\pm 3.9$ / $79.8\pm 0.6$
DebiAN	$45.4\pm 0.6$ / $52.6\pm 2.7$	$92.8\pm 1.1$ / $65.2\pm 8.9$	$80.0\pm 2.9$ / $79.6\pm 0.8$
GroupDRO	$40.7\pm 2.5$ / $64.1\pm 3.5$	$92.5\pm 0.1$ / $72.2\pm 2.7$	$81.3\pm 2.2$ / $84.4\pm 6.2$

Appendix

TABLE S6: Existing frozen-backbone probe results, mean

\pm

sample standard deviation over three seeds, percentages. Each pair is base / +BFR. Balanced WGA is included as a protocol check, not an optimal-head bound.

Explore similar work

Bias Leaves a Gradient Trail: Label-Free Bias Identification via Gradient Probes on Concept Decompositions
May 27, 2026Thomas Vitry, Kieran Edgeworth, Stefan Wermter +1Frozen Vision-Language Models In-Distribution
University of Hamburg, Hamburg, Germany · Ecole Normale Superieure de Rennes, Rennes, France
Dual-Branch Cross-Projection Debiasing through Diffusion-based Disentanglement
Jun 23, 2026Xiangqian Zhao, Xinyang Jiang, Zhipeng Xu +5Debiasing Unsupervised
Xidian University · Shaanxi, China · Microsoft Research Asia +1
Consistent Evidence, Robust Recognition: Faithful Attribution Regularization under Geometric Transformations
Jul 26, 2026Xianghao Jiao, Ruoyu Chen, Wei Wang +6Gradient-Weighted Class Activation Mapping Regularization
School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen, China · University of Chinese Academy of Sciences, Beijing, China · School of Automation, Nanjing University of Information Science and Technology +2

May 27, 2026, cs.CVJ/K move · Enter open · S save

Bias Leaves a Gradient Trail: Label-Free Bias Identification via Gradient Probes on Concept Decompositions

Thomas Vitry, Kieran Edgeworth, Stefan Wermter, Jae Hee Lee

University of Hamburg, Hamburg, Germany · Ecole Normale Superieure de Rennes, Rennes, France

Frozen Vision-Language Models In-Distribution Decomposition

Open paper AskPDF arXiv

Vision classifiers can exploit spurious correlations, achieving high in-distribution accuracy yet failing under distribution shift. Existing approaches to bias mitigation and analysis often depend on curated datasets, spurious-attribute or group labels, or retraining, which may be infeasible once a model is deployed or the relevant bias is unknown. We present a bias-label-free, post-hoc method for identifying spurious concepts in frozen vision models, relying only on standard class labels from a held-out audit dataset. For each target class, we collect patches from inputs predicted as that class and apply non-negative matrix factorization to intermediate activations to obtain a bank of interpretable concept vectors. Candidate concepts are then ranked with a bias estimator derived from their interaction with backpropagated gradients on misclassified examples: bias concepts tend to get activated when correcting false negatives and suppressed when correcting false positives. On Colored MNIST and Waterbirds the method recovers concepts aligned with the known spurious cue, and on CelebA it surfaces decision-relevant directions that only partially coincide with the annotated gender attribute; suppressing the top-ranked concepts at inference time improves worst-group accuracy by up to 17.9 percentage points on Waterbirds and 10.4 on CelebA without any retraining or parameter updates. Our method identifies decision-relevant spurious directions that need not coincide with annotated ones, providing both an interpretable auditing tool and an actionable debiasing handle for frozen vision models. Code is available at https://github.com/vitryt/label-free-bias-identification.

Abstract

Explore similar work

Bias Leaves a Gradient Trail: Label-Free Bias Identification via Gradient Probes on Concept Decompositions

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

Consistent Evidence, Robust Recognition: Faithful Attribution Regularization under Geometric Transformations