An Analysis of Random Projections in Cancelable Biometrics

With increasing concerns about security, the need for highly secure physical biometrics-based authentication systems utilizing \emph{cancelable biometric} technologies is on the rise. Because the problem of cancelable template generation deals with the trade-off between template security and matching performance, many state-of-the-art algorithms successful in generating high quality cancelable biometrics all have random projection as one of their early processing steps. This paper therefore presents a formal analysis of why random projections is an essential step in cancelable biometrics. By formally defining the notion of an \textit{Independent Subspace Structure} for datasets, it can be shown that random projection preserves the subspace structure of data vectors generated from a union of independent linear subspaces. The bound on the minimum number of random vectors required for this to hold is also derived and is shown to depend logarithmically on the number of data samples, not only in independent subspaces but in disjoint subspace settings as well. The theoretical analysis presented is supported in detail with empirical results on real-world face recognition datasets.

💡 Research Summary

The paper investigates why random projection (RP) is a crucial preprocessing step in cancelable biometric systems, providing both theoretical justification and empirical validation. Cancelable biometrics aim to protect biometric templates by applying a repeatable, reversible distortion so that compromised templates can be replaced without exposing the underlying physical traits. High‑dimensional feature vectors, common in modern biometric applications, incur significant computational and storage costs; thus dimensionality reduction is essential. However, conventional techniques such as PCA, LDA, or manifold learning often fail to preserve the subspace structure that is vital for discriminative template generation.

The authors formalize the notion of an “Independent Subspace Structure” (ISS) for a dataset. They consider K classes, each sampled from a linear subspace (S_i\subset\mathbb{R}^n). Subspaces are independent if no non‑zero vector in one can be expressed as a linear combination of vectors from the others, i.e., the direct sum (\bigoplus_{i=1}^K S_i) holds. The separation between any two subspaces is quantified by a margin (\gamma_{ij}), defined as the maximum inner product between unit vectors from the two subspaces (equivalently, the cosine of the principal angle). A dataset inherits this structure when each class’s samples lie in an independent subspace, and a class‑wise margin (\gamma_i) bounds the similarity between that class and all others.

Building on the Johnson‑Lindenstrauss (JL) lemma, Lemma 4 shows that a random matrix (R\in\mathbb{R}^{m\times n}) with i.i.d. Gaussian entries (or suitably sparse equivalents) preserves the Euclidean norm of any vector within a factor ((1\pm\epsilon)) with high probability. Lemma 5 extends this to cosine preservation: for any pair of vectors (x,y), the inner product after projection satisfies a bounded additive error that depends on (\epsilon) but not on the angle, guaranteeing that both acute and obtuse angles are approximately retained. The authors note that when the original cosine is near zero (nearly orthogonal vectors), the additive error can dominate, but this scenario corresponds to a very small margin, which is already favorable for classification.

The central theoretical contribution is Theorem 7, which addresses the multiclass case. It states that if a dataset has ISS with class margins (\gamma_i), then after projecting with a random matrix of dimension (m), each class’s new margin (\bar\gamma_i) satisfies (\bar\gamma_i \le (1-\epsilon)\gamma_i + \epsilon) with probability at least (1-6N^2\exp(-c m\epsilon^2)), where (N) is the total number of samples. Importantly, the required projection dimension scales logarithmically with (N) (i.e., (m = O(\log N))), meaning that a modest reduction in dimensionality suffices to preserve the subspace structure. The theorem also applies to disjoint but not necessarily independent subspaces, broadening its relevance.

To validate the theory, the authors conduct experiments on real‑world face recognition datasets. They generate random projection matrices of varying target dimensions (50 to 300) and apply them to high‑dimensional facial feature vectors. After projection, they employ a sparse representation‑based classifier (SR) that solves a basis‑pursuit problem using the projected training samples as an overcomplete dictionary. Recognition rates remain virtually unchanged across all reduced dimensions, confirming that discriminative power is retained. Additionally, they empirically estimate the probability of margin violation (rejection probability) for both acute and obtuse angles, showing close alignment with the bounds derived from Lemma 5 and Theorem 7. Figures illustrate that even for large negative cosines (obtuse angles), the cosine preservation holds, contradicting earlier claims in the literature.

The paper concludes that random projection serves as an effective “cancelable transform”: (1) it preserves the independent subspace structure essential for discriminative matching; (2) it maintains class margins, ensuring that inter‑class separability is not degraded; and (3) it dramatically reduces computational load while providing a straightforward mechanism to re‑issue new templates by simply drawing a fresh random matrix. The authors suggest future work on extending the analysis to non‑linear manifolds, handling non‑normalized features, and integrating RP into real‑time biometric pipelines. Overall, the study provides a rigorous foundation for the widespread adoption of random projection in secure, cancelable biometric systems.