Abstract
Deploying an underwater acoustic target recognition (UATR) model in a new water area combines two constraints the literature has treated separately: source data may be unavailable, motivating source-free unsupervised domain adaptation (SFUDA), and the water may hold vessel types never seen in training, motivating open-set recognition. We study their intersection — open-set source-free domain adaptation (OSFUDA) — and make the deployment reality the centerpiece: the agent receives the unlabeled target stream as-is, with naturally occurring unknown classes outnumbering knowns about three to one (nearly six in a third), no sampling, weighting, or filtering — the setting curated benchmarks abstract away. On three cross-water directions (Oceanship ↔ QiandaoEar22, Oceanship→VTUAD) under a shared two-class label space, reusing [1]'s source checkpoints, we evaluate three SFUDA mechanisms against source-only and supervised-oracle references over three seeds, jointly scoring known-class macro-F1 and unknown rejection (AUROC, FPR@95TPR, OSCR). Unknown contamination costs BN-statistics adaptation almost nothing (≤ 0.009 macro-F1 against the clean pools of [3]) while keeping its large recovery over source-only (+0.153 macro-F1); entropy minimization collapses reproducibly (2 of 3 seeds per direction), and contamination neither causes nor prevents it. Yet no post-hoc logit score (MSP, Energy, OpenMax) or source-prototype feature-space score yields a stable, calibratable, transferable rejection operating point: pooled AUROC stays near chance (adaptation-cell means 0.40–0.57), even the supervised oracle accepts 79–89% of unknowns at 95% known acceptance. Adaptation survives; post-hoc rejection fails. Unknown-aware representation learning and calibration — not further closed-set adaptation — are the central open need for deployable open-world UATR.