Split Learning (SL) offers a framework for collaborative model training that respects data privacy by allowing participants to share the same dataset while maintaining distinct feature sets. However, SL is susceptible to backdoor attacks, in which malicious clients subtly alter their embeddings to insert hidden triggers that compromise the final trained model. To address this vulnerability, we introduce SecureSplit, a defense mechanism tailored to SL. SecureSplit applies a dimensionality transformation strategy to accentuate subtle differences between benign and poisoned embeddings, facilitating their separation. With this enhanced distinction, we develop an adaptive filtering approach that uses a majority-based voting scheme to remove contaminated embeddings while preserving clean ones. Rigorous experiments across four datasets (CIFAR-10, MNIST, CINIC-10, and ImageNette), five backdoor attack scenarios, and seven alternative defenses confirm the effectiveness of SecureSplit under various challenging conditions.

The deployment of sensors enables data-driven urban management, but necessitates inductive spatio-temporal kriging to infer unmonitored areas. Existing methods impute these unknown observations by smoothing temporal features based on spatial dependencies, overlooking the decoupling ofinherent propertiesanddynamic correlationsin message passing. In particular, the inherent properties reveal non-transitive signals, and current coupled aggregation leads to inaccurate results. To this end, we proposeTempoRAl deCoupledKriging, named TRACK, to decouple two factors with the help of node-specific inherency. Specifically, we first construct a node-specific profile to represent its inherency including geographical and periodic features, which is subsequently transformed into decoupling prompts. Secondly, the coupled temporal features are separated through querying each prompt embedding, facilitating precise temporal aggregation for inherent properties and spatial aggregation for dynamic correlations. Finally, a multi-task training strategy is further adopted to mimic the inductive scenarios during testing. We evaluate TRACK on four real-world datasets spanning urban traffic and air quality prediction tasks. TRACK achieves state-of-the-art performance, with average improvements of 3.10% in MAE and 4.45% in RMSE over strong baselines. Moreover, we further demonstrated its robust generalization in a challenging cross-city inductive setting. Code is available athttps://github.com/JeremyChou28/TRACK.