Accurate skin lesion segmentation and classification remain challenging due to inter-class similarity, boundary ambiguity, class imbalance, and domain shifts across datasets. We introduce DualCoAg-Net, a novel multi-task deep learning architecture that integrates two core innovations: (i) a Contrastive Aggregation (CoAg) Block that aligns feature distributions across domains via supervised contrastive learning, improving class separation and generalization, and (ii) an Uncertainty-Aware Attention Decoder (UAD) that dynamically modulates feature fusion using pixel-level epistemic uncertainty, estimated efficiently through Bayesian Neural Networks. To handle data scarcity and imbalance, we incorporate conditional GAN (c-GAN) based augmentation for minority lesion synthesis. Adversarial training with PGD ensures robustness against perturbed or noisy clinical inputs. A federated learning framework with differential privacy enables training across five institutions without compromising patient confidentiality. Multi-task outputs are optimized using Focal Tversky Loss (segmentation), class-weighted cross-entropy (classification), and KL divergence (uncertainty regularization). Segmentation boundaries are refined using Conditional Random Fields (CRFs) to achieve high accuracy. Evaluated on ISIC 2018, HAM10000, and Derm7pt, the propose model achieves segmentation Dice: 93.1%, IoU: 87.6%, Accuracy: 97.2%, and AUC: 98.4%, outperforming STM-UNet and MDSA-UNet like segmentation baselines. Similarly in classification, DualCoAg-Net 91.7% accuracy, 90.9% F1-score, and AUC 0.948, compared to TriST and DSCC_Net like counterparts. Cross-Domain evaluations using DualCoAg-Net with Leave-One-Dataset-Out principle on three datasets proven the generalization capabilities of the proposed framework. Ablation studies validated the importance of the DualCoAg-Net modules and their effect on various segmentation and classification metrics. This framework bridges critical gaps in robustness, interpretability, and generalizability, offering a clinically viable and technically advanced solution for comprehensive skin lesion analysis.