Command-and-Control (C C) communications are challenging to detect because low-volume, periodic, and adaptive traffic can resemble legitimate network activity. This paper presents an explainable behavioral flow abstraction framework for C C detection in IoT and cloud environments using observable packet-, flow-, temporal-, directional-, and transport-level characteristics without payload inspection. The framework constructs a compact behavioral representation and refines it through correlation analysis, Variance Inflation Factor (VIF), Recursive Feature Elimination (RFE), and Denoising Autoencoder (DAE)-assisted feature refinement, resulting in a selected 16-feature representation. Evaluation includes binary and multiclass classification, ablation, repeated cross-validation, leave-one-scenario-out (LOSO) validation, explainability analysis, benign-capture holdout evaluation, and independent Azure validation. Under within-benchmark flow-level evaluation, LightGBM achieved 99.79
Underwater images are degraded by factors such as light absorption and scattering, which increases the difficulty of downstream visual tasks. However, most existing underwater image restoration methods rely on limited receptive fields to extract features in the original pixel/spatial domain, leading to suboptimal performance in restoring high-quality underwater images. To address this, we propose a Receptive-Field-Expanded Spatial-Frequency-Aware Cascaded Network (RSFNet) for underwater image restoration (UIR). The network integrates multiple feature enhancement blocks (FEBs) at each encoder-decoder scale. Each FEB consists of a spatial domain enhancement module and a multi-branch frequency-aware decomposition module. The spatial domain enhancement module expands the receptive field using multi-scale dilated convolutions and asymmetric convolutions, capturing a broader contextual information. The frequency-aware enhancement module further strengthens the modeling of frequency-aware features through a directional strip pooling strategy, combining both local and global receptive fields. This deep fusion of spatial and frequency-aware information effectively improves image restoration quality and enhances the network’s adaptability to underwater image degradation. Extensive experimental results on the UIEBD, LSUI, and U45 benchmark datasets demonstrate that RSFNet outperforms existing underwater image restoration methods in metrics such as PSNR, SSIM, LPIPS, and FID, achieving a good balance between model size and performance.
Long-context inference on memory-constrained graphics processing unit (GPU) systems depends on configured capacity, runtime placement, and requested output. We present LC-Select, a profile-guided routing method for latency, throughput, GPU energy, and stability. We profile Q4_K_M Gemma 4 E2B under Ollama 0.24.0 on RTX 3060 8 GB and RTX A2000 12 GB systems, freeze the profiles, and evaluate a separate workload with new prompts and output budgets using three independent requests per condition. At a configured capacity of 65,536 tokens, the RTX 3060 runtime reports mixed host/device model weight placement and lower decode throughput. LC-Select matches all 20 single-objective and 20 weighted held-out oracle decisions; a 264-decision weight grid contains one low-regret mismatch. Output-aware estimation reduces mean absolute percentage error for GPU energy per output token from 17.75 to 2.36
Evaluating long-horizon robot manipulation policies is a high-throughput computing workload: one CALVIN-style campaign sweeps checkpoints, seeds and task sequences across many concurrent simulator workers, consuming tens of thousands of rollouts before any score is inspected. Evaluation-path correctness is therefore a workflow-validation concern in parallel evaluation: an adapter defect shared by workers can yield a well-formed table of zeros while scheduler and determinism checks pass. We report a CALVIN case study and use it to propose the Benchmark-Fidelity Diagnostic Protocol (BFDP): a cheap-first set of single-worker controls, run before a parallel sweep, that audits export horizon, teacher replay semantics and reset fidelity in a fixed order. On Composing Actions from Language and Vision (CALVIN) scene D with a vision-language-to-diffusion pipeline, BFDP separates three findings of three different kinds. For the two tasks with a measured effect onset, the fixed 8-step export ends before the observed onset, and across benchmark-critical tasks it retains only 12–16
Accurate prediction of dam displacement is essential for assessing long-term service safety and supporting dam operation and management. Existing deep learning approaches, particularly graph-based models, have demonstrated potential in multi-point deformation prediction. However, many existing approaches rely on predefined or fixed spatial relationships, which may not adequately characterize directional interactions among monitoring points or distinguish the varying importance of neighboring nodes. To address these challenges, this study proposes a dam displacement prediction framework (GD-GAT) that integrates spatial priors with a gated graph attention network. The framework first employs a multi-scale long short-term memory (LSTM) network to extract temporal features from historical displacement and water level series and incorporates time-scale attention to adaptively fuse multi-scale temporal information. Subsequently, a dual-path graph attention module is designed to adaptively learn asymmetric and heterogeneous spatial dependencies among monitoring points. Furthermore, a coordinate-based spatial prior constructed using a Gaussian kernel is incorporated through a gated fusion mechanism, enabling the model to jointly exploit prior spatial relationships and data-driven dependencies. Experimental results on a real-world dam monitoring dataset demonstrate that, compared with the best-performing baseline, GD-GAT reduces RMSE and MAE by 13.94
