Mobile Edge Computing (MEC) facilitates the rapid response and energy-efficient execution of tasks on mobile devices. However, determining whether and where to offload tasks remains a significant challenge due to the constantly changing character of workloads in MEC environments. To address this issue, this paper proposes PreAlloc-A2C—a deep reinforcement learning actor-critic-based framework that calculates allocation scores by leveraging both task features (task size, required completion time, and waiting time) and server features (queue length and historical workload). This design enables fully distributed task offloading decisions without centralized coordination. Additionally, a Long Short-Term Memory (LSTM) network is integrated to forecast impending server loads, thereby supporting adaptive scheduling. A tailored reward function is also designed to jointly optimize three key performance metrics: task delay, device energy consumption, and task drop rate. Extensive experiments are conducted to evaluate PreAlloc-A2C against five baseline algorithms: Particle Swarm Optimization (PSO), Advantage Actor-Critic (A2C), Deep Q-Network (DQN), Double Deep Q-Network (DDQN), and Dueling Deep Q-Network (Dueling DQN). The results show that PreAlloc-A2C outperforms all baselines, achieving lower latency, reduced energy consumption, and a lower task drop rate.

Graph neural networks (GNNs) have demonstrated impressive capabilities in processing graph-structured data, yet their vulnerability to adversarial perturbations poses serious challenges to real-world applications. Existing defense methods often fail to handle diverse types of attacks and adapt to dynamic adversarial strategies because they typically rely on static defense mechanisms or focus narrowly on a single robustness dimension. To address these limitations, we propose an adversarial attention-based robustness strategy (AARS), which is a unified framework designed to enhance the robustness of GNNs against structural and feature perturbations. AARS operates in two stages: the first stage employs adversarial training with joint optimization to improve the resilience of the model to malicious attacks and stabilize its decision boundaries; the second stage incorporates an attention mechanism to identify critical structural dependencies and guides a topology filtering module that dynamically suppresses adversarial edges while preserving essential graph semantics. Extensive experiments on benchmark datasets for node classification demonstrate that AARS significantly outperforms existing baselines in terms of classification accuracy under various attack scenarios, thereby effectively improving the robustness of GNNs in static and dynamic adversarial settings.

Electrocardiogram (ECG) is a widely used non-invasive tool for diagnosing cardiovascular diseases. ECG zero-shot classification involves pre-training a model on a large dataset to classify unknown disease categories. However, existing ECG feature extraction networks often neglect key lead signals and spatial topology dependencies during cross-modal alignment. To address these issues, we propose a multimodal channel compression graph attention alignment network (MCCGAA). MCCGAA incorporates a channel attention module (CAM) to effectively integrate key lead features and a graph attention-based alignment network to capture spatial dependencies, enhancing cross-modal alignment. Additionally, MCCGAA employs a log-sum-exp loss function, improving classification performance and convergence over the original clip-style method. Experimental results show that MCCGAA outperforms current methods, achieving the highest classification accuracy across six publicly available datasets. MCCGAA holds promise for advancing ECG zero-shot classification and offering better decision support for researchers.