Fine-tuning large pre-trained foundation models often yields excellent downstream performance but is prohibitively expensive when updating all parameters. Parameter-efficient fine-tuning (PEFT) methods such as LoRA alleviate this by introducing lightweight update modules, yet they commonly rely on weight-agnostic linear approximations, limiting their expressiveness. In this work, we propose PEANuT, a novel PEFT framework that introduces weight-aware neural tweakers, compact neural modules that generate task-adaptive updates conditioned on frozen pre-trained weights. PEANuT provides a flexible yet efficient way to capture complex update patterns without full model tuning. We theoretically show that PEANuT achieves equivalent or greater expressivity than existing linear PEFT methods with comparable or fewer parameters. Extensive experiments across four benchmarks with over twenty datasets demonstrate that PEANuT consistently outperforms strong baselines in both NLP and vision tasks, while maintaining low computational overhead.

Accurate localization is fundamental for internet of things (IoT) applications but introduces significant privacy risks in crowdsourcing-based systems. Conventional privacy-preserving techniques suffer from critical limitations: subtraction-based models sacrifice positioning accuracy by discarding measurement equations; cryptographic solutions incur prohibitive computation overhead; while secret-sharing approaches expose target locations. To address these challenges, we propose the mean-based subtraction localization (MSL) model that preserves all measurement equations through global averaging, eliminating information loss while establishing a privacy-conducive mathematical framework. Building on MSL, we develop the mean-based privacy-preserving localization (MPPL) algorithm featuring: 1) matrix decomposition for bidirectional privacy isolation, 2) lightweight secure summation/multiplication primitives, and 3) linear-complexity operations. Rigorous theoretical verification confirms MPPL’s feasibility, correctness, information-theoretic privacy, and efficiency. Experimental results demonstrate 10% – 14% higher accuracy than privacy-preserving multi-lateral localization (PPPL)/efficient privacy-preserving localization (EPPL) under noise while reducing computation time by a factor of 2.8 versus homomorphic encryption schemes, establishing a new paradigm for high-precision privacy-aware localization.

Accurate typhoon track forecasting is crucial for early system warning and disaster response. While Transformer-based models have demonstrated strong performance in modeling the temporal dynamics of dense trajectories of humans and vehicles in smart cities, they usually lack access to broader contextual knowledge that enhances the forecasting reliability of sparse meteorological trajectories, such as typhoon tracks. To address this challenge, we propose TyphoFormer, a novel framework that incorporates natural language descriptions as auxiliary prompts to improve typhoon trajectory forecasting. For each time step, we use Large Language Model (LLM) to generate concise textual descriptions based on the numerical attributes recorded in the North Atlantic hurricane database. The language descriptions capture high-level meteorological semantics and are embedded as auxiliary special tokens prepended to the numerical time series input. By integrating both textual and sequential information within a unified Transformer encoder, TyphoFormer enables the model to leverage contextual cues that are otherwise inaccessible through numerical features alone. Extensive experiments are conducted on HURDAT2 benchmark, results show that our TyphoFormer consistently outperforms other state-of-the-art baseline methods, particularly under challenging scenarios involving nonlinear path shifts and limited historical observations.

The rapid proliferation of user terminals and bandwidth-intensive applications has posed significant challenges to traditional centralized network architectures, thereby underscoring the necessity of the mobile edge caching (MEC) technique, which brings frequently requested data closer to end users. However, due to the inherently limited cache capacity at edge nodes, the effectiveness of MEC is highly dependent on the cache update strategies. Conventional caching strategies adapt poorly to dynamic content popularity, whereas reinforcement learning (RL) methods are often sensitive to hyperparameter choices and prone to becoming trapped in local optima. To overcome these limitations, this paper proposes a novel cache content update strategy based on the teaching-learning-based optimization (TLBO) algorithm, referred to as the MECTLBO strategy. The proposed approach is characterized by low parameter sensitivity and strong global optimization capability. Extensive simulation results demonstrate that the MECTLBO strategy outperforms conventional and RL-based caching strategies, exhibiting robust adaptability to dynamic content popularity patterns, varying cache sizes, and diverse content volumes, as well as consistent effectiveness across multiple test datasets.

Model extraction attacks pose significant security threats to deployed language models, potentially compromising intellectual property and user privacy. This survey provides a comprehensive taxonomy of LLM-specific extraction attacks and defenses, categorizing attacks into functionality extraction, training data extraction, and prompt-targeted attacks. We analyze various attack methodologies including API-based knowledge distillation, direct querying, parameter recovery, and prompt stealing techniques that exploit transformer architectures. We then examine defense mechanisms organized into model protection, data privacy protection, and prompt-targeted strategies, evaluating their effectiveness across different deployment scenarios. We propose specialized metrics for evaluating both attack effectiveness and defense performance, addressing the specific challenges of generative language models. Through our analysis, we identify critical limitations in current approaches and propose promising research directions, including integrated attack methodologies and adaptive defense mechanisms that balance security with model utility. This work serves NLP researchers, ML engineers, and security professionals seeking to protect language models in production environments.

Efficiently capturing the complex spatiotemporal representations from large-scale traffic data with uneven data quality remains to be a challenging task. In considering of the dilemma, this work employs the advanced contrastive learning and proposes a novel Spatial-Temporal Synchronous Contextual Contrastive Learning (STS-CCL) model. First, we elaborate the basic and strong augmentation methods for spatiotemporal graph data. Second, we introduce a Spatial-Temporal Synchronous Contrastive Module (STS-CM) to simultaneously capture the decent spatial-temporal dependencies and realize graph-level contrasting. To further discriminate node individuals in negative filtering, a Semantic Contextual Contrastive method is designed based on semantic features and spatial heterogeneity, achieving node-level contrastive learning along with negative filtering. Finally, we present a hard mutual-view contrastive training scheme and extend the classic contrastive loss to an integrated objective function, yielding better performance. Extensive experiments and evaluations demonstrate that building a predictor upon STS-CCL contrastive learning model gains superior performance than existing traffic forecasting benchmarks. The proposed STS-CCL is highly suitable for large datasets with only a few labeled data and other spatiotemporal tasks with data scarcity issue.

Accurate citywide traffic flow prediction is of great importance to intelligent transportation system. Existing methods typically assume the complete citywide traffic data can be obtained in real-time, which is impossible in applications. Furthermore, many recent works only consider one single kind of spatial correlation in traffic network when building graph representations. This work proposes an adversarial learning framework named Multi-Graph Convolutional Generative Adversarial Networks (MGC-GAN) to address the aforementioned challenges. To generate citywide traffic flow predictions using limited traffic data, we construct three kinds of graphs using easily accessed geographical and semantic information to model the complex spatial correlations in citywide transportation networks. Following that, a parallel GCN layer is designed to separately process multiple graphs. In addition, we design the Parallel Graph Convolution and Temporal Convolution Module (PGTCM) to effectively capture the heterogeneous spatial-temporal dependencies. Extensive experiments are carried out on two citywide traffic datasets, demonstrating that MGC-GAN outperforms several state-of-the-art baseline methods.

Mobile edge caching is an emerging approach to manage high mobile data traffic in fifth‐generation wireless networks that reduces content access latency and offloading data traffic of backhaul links. This paper proposes a novel cooperative caching policy based on long short‐term memory (LSTM) neural networks considering the characteristics between the features of the heterogeneous layers and the user moving speed. Specifically, LSTM is applied to predict content popularity. Size‐weighted content popularity is utilised to balance the impact of the predicted content popularity and content size. We also consider the moving speeds of mobile users and introduce a two‐level caching architecture consisting of several small base stations (SBSs) and macro base stations (MBSs). To avoid content requests of fast‐moving users affecting the content popularity distribution of the SBS since fast‐moving users frequently handover among SBSs, fast‐moving users are served by MBSs no matter which SBS they are in. SBSs serve low‐speed users, and SBSs in the same cluster can communicate with one another. The simulation results show that compared to common cache methods, for example, the least frequently used and least recently used methods, our proposed policy is at least 8.9% lower and 6.8% higher in terms of the average content access latency and offloading ratio, respectively.

This paper proposes a DRL-based cache content update policy in the cache-enabled network to improve the cache hit ratio and reduce the average latency. In contrast to the existing policies, a more practical cache scenario is considered in this work, in which the content requests vary by both time and location. Considering the constraint of the limited cache capacity, the dynamic content update problem is modeled as a Markov decision process (MDP). Besides that, the deep Q-learning network (DQN) algorithm is utilised to solve the MDP problem. Specifically, the neural network is optimised to approximate the Q value where the training data are chosen from the experience replay memory. The DQN agent derives the optimal policy for the cache decision. Compared with the existing policies, the simulation results show that our proposed policy is 56%–64% improved in terms of the cache hit ratio and 56%–59% decreased in terms of the average latency.