As image data are shared ever more openly, they increasingly leak privacy-sensitive information, yet existing detectors seldom model how the text embedded in an image relates to its visual content and tend to fail precisely when the two modalities disagree. We note that the conventional softmax attention used for multimodal fusion carries an implicit sign symmetry: Every source token contributes only additively, so conflicting evidence is averaged away rather than resolved. We propose a symmetric dual-source fusion framework whose decoder deliberately breaks this sign symmetry. Image and text are first encoded by a Swin Transformer and a policy knowledge-enhanced BERT (KL-BERT) and projected into a common space to form a permutation-symmetric dual-source memory. A Signed Cross-attention Auto-compressing Decoder (SCAD) then fuses the two sources through an attention map that factorizes into a sign-symmetric (even) magnitude term and a sign-antisymmetric (odd) polarity term, allowing the model to either reinforce or actively subtract cross-modal evidence. Experiments on a self-constructed privacy-sensitive image dataset show that the proposed method attains an accuracy of $\mathrm{97.69\%}$ and an F1-score of $\mathrm{96.70\%}$, with its largest gains on the face category, the most conflict-prone class in our data, suggesting that controlled symmetry breaking is an effective principle for cross-modal fusion.

Missing-data strategies are often compared using a single cohort or prediction model, making it difficult to distinguish imputation-specific effects from benefits due to the downstream learner or prediction averaging. We conducted a leakage-controlled factorial simulation in two public cardiac cohorts (n = 299 and n = 918), comparing six strategies across three missingness mechanisms, three missingness rates (10–50%), ten seeds, and three learners. A four-arm analysis further examined the contribution of multiple imputation relative to single imputation and model averaging. Strategy effects depended on the learner: exploratory strategy-by-learner interactions were significant in both cohorts, while no strategy effect survived Holm correction for penalised logistic regression. Prediction-layer-pooled multiple imputation ranked first for the tree-based learners in the larger cohort. In the smaller cohort with XGBoost, its paired median AUROC difference versus a bagged mean-imputation ensemble was −0.003 (Holm-adjusted p = 0.275); because equivalence was not tested, this result does not establish that the strategies perform equivalently. No consistent strategy-by-missingness-mechanism interaction was found. These findings support evaluating missing-data handling jointly with the prediction model and interpreting apparent strategy advantages in light of ensemble effects. This research offers new insights into strategies for handling missing data.