Conventional waste monitoring relies heavily on manual inspection, while most detection models are trained on close-range, simplified datasets, limiting their applicability for real-world surveillance. Even with surveillance imagery, challenges such as cluttered backgrounds, scale variation, and small object sizes often lead to missed detections and reduced robustness. To address these challenges, this study introduces EcoDetect-YOLOv2, a lightweight and high-efficiency object detection model developed using the Intricate Environment Waste Exposure Detection (IEWED) dataset. Building upon the YOLOv8s architecture, EcoDetect-YOLOv2 incorporates a small object detection P2 detection layer to enhance sensitivity to small objects. The integration of an efficient multi-scale attention (EMA) mechanism prior to the P2 head further improves the model’s capacity to detect small-scale targets, while bolstering robustness against cluttered backgrounds and environmental noise, as well as generalizability across scale variations. In the feature fusion stage, a Dynamic Upsampling Module (Dysample) replaces traditional nearest-neighbor upsampling to yield higher-quality feature maps, thereby facilitating improved discrimination of overlapping and degraded waste particles. To reduce computational overhead and inference latency without sacrificing detection accuracy, Ghost Convolution (GhostConv) replaces conventional convolution layers within the neck. Based on this, a GhostResBottleneck structure is proposed, along with a novel ResGhostCSP module—designed via a one-shot aggregation strategy—to replace the original C2f module. Experiments conducted on the IEWED dataset, which features multi-object, multi-class, and highly complex real-world scenes, demonstrate that EcoDetect-YOLOv2 outperforms the baseline YOLOv8s by 1.0%, 4.6%, 4.8%, and 3.1% in precision, recall, mAP₅₀, and mAP_50:95, respectively, while reducing the parameter count by 19.3%. These results highlight the model’s effectiveness in real-time, multi-object waste detection, providing a scalable and efficient tool for automated urban and digital governance.

Prompt injection detection is commonly studied as a static offline classification problem, yet deployed LLM systems face evolving attacks and distribution shift after deployment. Static detectors are therefore poorly matched to the threat model, while routing every input to a stronger external LLM is costly and defeats the purpose of a local detector. We formulate prompt injection detection as a selective test-time adaptation problem. Our framework combines a prompt-based local detector built on masked language modeling and a learnable soft verbalizer with an entropy-based active querying mechanism that escalates only high-uncertainty inputs to an external LLM. Queried hard samples are then stored in a review window and replayed for subsequent detector updates. Empirical evaluations across multiple benchmarks show that EvoShield achieves performance on par with or even exceeding pure Large Language Model baselines, while cutting API query costs by more than 85%.