Real-world IoT systems often require collaborative learning across multiple tasks performed by heterogeneous devices with varying computational resources. Conventional multi-task federated learning typically shares a pre-trained encoder across clients and trains task-specific classifiers locally. However, freezing the encoder during training leads to over-reliance on pre-training data and limits the model’s ability to capture inter-task relationships, reducing multi-task learning accuracy. Furthermore, significant variations in device capabilities highlight the need for heterogeneous model adaptation, making the unified model architecture assumption impractical for real-world IoT scenarios. To address these challenges, we propose MTFedSD, a novel multi-task federated learning across heterogeneous systems via self-distillation method. MTFedSD introduces multiple models of varying sizes and architectures to match clients’ resource capabilities, while enabling selective parameter sharing to extract consistent and transferable representations across tasks. A heterogeneous model aggregation strategy and self-distillation mechanism are integrated to reconcile differences across tasks and models, mitigating conflicts during joint training. Furthermore, the joint optimization strategy of shared encoder and pruning improves the communication efficiency in multi-task scenarios. Moreover, we provide a theoretical analysis that characterizes the convergence behavior of MTFedSD under heterogeneous and multi-task settings. Extensive experiments show that MTFedSD achieves robust and scalable performance across diverse tasks and device types, offering a practical solution for federated learning in complex IoT environments.

Nowadays, most edge devices running Android still adopt the memory reclaim scheme designed for the server, which easily leads to a large number of page re-faults in the kernel, making mobile device workloads inefficient. We hold the opinion that the amount of reclaimable pages is sequential and predictable. Therefore, we propose an adaptive lightweight memory management scheme based on prediction, which consists of two parts: a lightweight and optimal prediction model generation module (LOPMG) and an adaptive prediction-based reclaim scheme (APRS). LOPMG adopts the Particle Swarm Optimization (PSO) algorithm to generate and quantify Long-Short Term Memory (LSTM) prediction model. By APRS, the PSO-LSTM model works in the kernel and can predict allocation workloads to tune reclaim parameters dynamically. Experiments show that our PSO-LSTM model can improve the accuracy of prediction, and the reclaim scheme can significantly reduce the number of page re-faults and the time of application relaunch.

With the rapid development of wireless communication and localization technologies, the easier collection of trajectory data can bring potential data-driven value. Recently, there has been an increasing interest in how to publish trajectory dataset without revealing personal information. However, since the large-scale and real-world sequential trajectory dataset presents a heterogeneous regional distribution, the existing study ignores the relationship between privacy budget allocation and spatial characteristics, resulting in unreasonable continuity and mapping distortion, and thus lowering the utility of the synthetic dataset. To address this problem, we propose a probability distribution model named Adaptive grid-based Weighted Differential Privacy (AWDP). First, trajectories are adaptively discretized into the multi-resolution grid structures to make trajectories more uniformly distributed and less disturbed by the noise. Second, we allocate different weighted budgets for different grids according to density-based regional characteristics. Third, a spatio-temporal continuity maintenance method is designed to solve unrealistic direction- and density-based continuity deviations of synthetic trajectories. An application system is developed for demonstration purposes which is available online athttp://qgailab.com/awdp/. The extensive experiments on three datasets demonstrate that AWDP performs significantly better than the state-of-the-art model in preserving the density distribution of the original trajectories with differential privacy guarantee and high utility.