Towards Identity Unlearning for Face Verification: Distribution Alignment and Feature Dispersion

Face verification infers identity through pairwise embedding similarity, yet most machine unlearning methods target classification and assess success by suppressing target-class predictions. This criterion is inadequate for face verification: a sample can be misclassified while same-identity embeddings remain compact and separable, leaving the individual reliably verifiable. We formulate face verification unlearning, requiring retained-identity verification to be preserved while accuracy approaches chance for both forgotten-forgotten and forgotten-retained pairs. This reveals two failure modes: residual within-identity compactness and separation from the retained population. We therefore recast unlearning as assimilation and show that matching a forgotten identity's embedding distribution to the retained population is sufficient to make the two pair distributions indistinguishable. Guided by this result, we combine cosine-cost optimal transport for distribution alignment with identity-wise dispersion for breaking within-identity compactness. Our formulation explicitly targets the pairwise verifier by requiring separate chance-level behavior for forgotten-forgotten and forgotten-retained pairs. Experiments on three benchmarks show that our method moves verification involving forgotten identities toward chance while largely preserving retained-identity performance.