Federated learning (FL) offers a decentralized, privacy-preserving framework for encrypted traffic classification (ETC), enabling network management and security. However, real-world deployment of federated ETC faces compound client specific feature, concept, and label drift, which degrades model performance. Existing ETC methods under FL settings typically address these drift types in isolation or partial combinations, overlooking their entanglement. Moreover, multiple-global model and personalized FL approaches are computational and communication expensive. To fill this gap, we propose DAIR FedMoE, a Drift-Adaptive, Imbalance-Aware, RL-Managed Federated Mixture-of-Experts framework to simultaneously handle the drift triad with single-global model while minimizing the computational and communication overhead. DAIR-FedMoE in tegrates a GShard Transformer with a hierarchical Mixture of-Experts (MoE) layer that routes encrypsted flows to either stable or drift-specialist experts based on per-client drift scores. Within each expert, entropy-guided loss reweighting empha sizes low-confidence classes to address dynamic label imbalance. Additionally, a reinforcement learning-based policy dynamically manages the expert pool by spawning, pruning, and merging experts, enabling efficient adaptation to evolving traffic patterns. Experiments on federated splits of ISCX-VPN, ISCX-Tor, VNAT, and USTC-TFC2016 show that DAIR-FedMoE achieves superior macro-F1, minority-class recall, and drift-recovery speed compared to state-of-the-art baselines, while preserving privacy and communication efficiency. The source code is available at https://github.com/dairfedmoe/DairFM.

Generating realistic two-person interaction motions from text holds immense potential in computer vision and animations. While existing latent motion diffusion models offer compact and efficient representations, they often fail to produce physically plausible contacts and are typically constrained to a single canonical body shape. As a result, the generated motion sequences exhibit substantial mesh penetrations and lack interaction realism. To address these limitations, we propose a contact and shape-aware latent motion representation and diffusion model (CoShMDM) for generating realistic two-person interactions from text. Our framework begins by constructing contact-compatible motion using SMPL-based meshes and a normal alignment-based mesh contact matrix to capture fine-grained mesh-level contacts. To account for shape diversity, we incorporate SMPL shape parameters and iteratively learn contact dynamics across different body shapes. Additionally, a reinforcement learning-based mesh penetration avoidance policy network, guided by signed distance fields, is introduced to minimize mesh penetrations while preserving contact fidelity and shape-aware motion. We further employ a dual-encoder VQ-VAE to learn disentangled latent representations for motion and contacts, which are then utilized in a text- and body-shape-conditioned diffusion model. To ensure spatial, temporal, and semantic coherence, we integrate a novel contact and motion consistency module into the diffusion transformer. Extensive evaluations on the InterHuman and InterX datasets demonstrate that our method outperforms state-of-the-art approaches achieving lowest FID scores (4.801 and 0.013), with 19% and 17.3% reductions in mesh penetrations, and 17.8% and 33.2% gains in contact similarity, respectively.