With the rapid proliferation of artificial intelligence (AI) applications, an increasing number of edge devices—such as smartphones, cameras, and embedded controllers—are being tasked with performing AI-based inference. Due to constraints in storage capacity, computational power, and network connectivity, these devices are often categorized as operating in resource-constrained environments. In such scenarios, deploying powerful Transformer-based models like ChatGPT and Vision Transformers is highly impractical because of their large parameter sizes and intensive computational requirements. While lightweight Transformer models, such as MobileViT, offer a promising solution to meet storage and computational limitations, their robustness remains insufficient. This poses a significant security risk for AI applications, particularly in critical edge environments. To address this challenge, our research focuses on enhancing the robustness of lightweight Transformer models under resource-constrained conditions. First, we propose a comprehensive robustness evaluation framework tailored for lightweight Transformer inference. This framework assesses model robustness across three key dimensions: noise robustness, distributional robustness, and adversarial robustness. It further investigates how model size and hardware limitations affect robustness, thereby providing valuable insights for robustness-aware model design. Second, we introduce a novel adversarial robustness enhancement strategy that integrates lightweight modeling techniques. This approach leverages methods such as gradient clipping and layer-wise unfreezing, as well as decision boundary optimization techniques like TRADES and SMART. Together, these strategies effectively address challenges related to training instability and decision boundary smoothness, significantly improving model robustness. Finally, we deploy the robust lightweight Transformer models in real-world resource-constrained environments and empirically validate their inference robustness. The results confirm the effectiveness of our proposed methods in enhancing the robustness and reliability of lightweight Transformers for edge AI applications.

High-density cities face increasing pressure to curb the rapid growth of building-sector carbon emissions, yet the specific impacts of China’s dual-control policy—regulating both total emissions and emission intensity—remain insufficiently understood. To address this gap, this study aims to quantify how dual-control constraints shape urban building-emission trajectories by developing a dynamic scenario model that integrates both operational and embodied emissions while accounting for technological progress, energy-structure adjustments, and socioeconomic change. The model is applied to Guangzhou, a representative high-density metropolis in China’s low-carbon transition. Three policy scenarios are evaluated: a baseline pathway reflecting existing trends, an energy-conservation pathway emphasizing efficiency improvements, and a deep-decarbonization pathway that combines enhanced efficiency with accelerated clean-energy adoption. The results show that rising residential energy demand is the primary driver of emission growth, whereas technological advancement provides the strongest mitigation potential. Under the energy-conservation scenario, emissions are projected to peak around 2029, consistent with China’s national carbon-peaking target for 2030. Overall, the findings offer clear empirical evidence and actionable policy insights for accelerating urban decarbonization in high-density contexts under dual-control constraints.

In the planning and decision making of transport systems, data relationships are complex and the eigenvalues of data samples may contain errors, resulting in traffic fuzzy data. Traffic fuzzy data has seriously affected the model fitting and lagged the model performance. Optimising the traffic fuzzy data and extracting potentially usable information from it to assist in supporting the decision-making process has become an urgent problem to be solved. In order to solve the data ambiguity problem in the modelling process, this paper improves the structure of denoising diffusion probabilistic model (DDPM), deepens the depth and width of the model, and proposes a Time-Aware Multi-Head Attention Denoising Diffusion Probability Model (TAMHA-DDPM). By integrating the attention mechanism, residual block and affine transformation, replacing the original attention mechanism with a multiple attention mechanism, and using the residual block for up-sampling and down-sampling, the model’s recognition of fuzzy data is greatly improved. Comparison experiments on the parameters of the proposed algorithm on the KITTI dataset show that different degrees of fuzzy data structure reconstruction can be achieved by parameter adjustment, which is also reflected in the evaluation index. When the number of noise-adding rounds is about 200, the data structure is richer, and the performance of related indexes is better; when the number of denoising rounds is about 1000, the data structure is more complete, and the qualitative results are more close to the original data. The time-aware multi-attention denoising diffusion probabilistic model outperforms the existing models in data reconstruction performance, can effectively optimise the fuzzy data in intelligent transportation systems, and the recovered data structure results are more in line with the original data.