SOMnibus: Recovering Underlying Sensitive Attributes with Self-Organizing Maps

Unsupervised representation learning is often assumed to be benign with respect to sensitive attributes when those attributes are withheld from training. We challenge this assumption by demonstrating systematic \emph{representation-level leakage} of ordinal sensitive attributes in purely unsupervised embeddings. Using \textbf{SOMnibus}, a topology-preserving method based on high-capacity Self-Organizing Maps, we show that attributes such as age and income emerge as dominant latent axes despite being explicitly excluded from the input. Across two large-scale real-world benchmarks, the World Values Survey and the Census-Income (KDD) dataset, SOMnibusrecovers monotonic orderings aligned with withheld sensitive attributes, achieving Spearman correlations of up to $0.85$, while PCA and UMAP typically remain below $0.23$ . Moreover, unsupervised segmentation of SOMnibus embeddings yields demographically skewed clusters, revealing downstream fairness risks in the absence of any supervised task. These results demonstrate that \emph{fairness through unawareness} can fail at the representation level.