Efficient inference for on-device Large Language Models (LLMs) remains challenging due to limited hardware resources and the high cost of the prefill stage, which processes the full input context to construct Key-Value (KV) caches. We present SparKV, an adaptive KV loading framework that combines cloud-based KV streaming with on-device computation. SparKV models the cost of individual KV chunks and decides whether each chunk should be streamed or computed locally, while overlapping the two execution paths to reduce latency. To handle fluctuations in wireless connectivity and edge resource availability, SparKV further refines offline-generated schedules at runtime to rebalance communication and computation costs. Experiments across diverse datasets, LLMs, and edge devices show that SparKV reduces Time-to-First-Token by 1.3×-5.1× with negligible impact on response quality, while lowering per-request energy consumption by 1.5× to 3.3×, demonstrating its robustness and practicality for real-world on-device deployment.

Intracortical brain-computer interfaces (iBCIs) promise to sense brain activity at an unprecedented scale and resolution. However, unlocking this potential for practical, untethered applications remains an unsolved challenge. The major barrier is the significant wireless bandwidth required to stream high-resolution brain signals. Existing approaches rely on extensive on-device processing, which is severely constrained by the limited resources of iBCI devices, the complexity of brain signals, and the dynamic nature of neural activity. This paper introduces Neuralite, a wireless iBCI system that integrates high-fidelity brain signal models and effective brain sensing mechanisms within an efficient server-driven streaming framework. By thoroughly characterizing brain signal variability, Neuralite adaptively optimizes streaming under dynamic neural conditions, minimizing bandwidth consumption without imposing excessive burdens on resource-constrained iBCI devices. Experimental results demonstrate that Neuralite significantly reduces bandwidth consumption while preserving neural decoding precision across key iBCI components and representative applications.

With the commercialization of fifth-generation (5G), the rapid popularity of mobile Over- The- Top (OTT) voice applications brings huge impacts on the traditional telecommunications voice call service. Tunnel encryption and anonymous network technologies allow OTT users to escape the supervision of network operators easily, which may cause potential security risks to cyberspace. To monitor harmful OTT applications in the context of 5G, it is critical to identify encrypted OTT voice traffic. However, there is no comprehensive study on typical OTT voice traffic identification. This is the first study to analyze OTT Virtual Private Network (VPN) voice traffic in the 5G network specifically. We propose to employ Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) to classify encrypted 5G OTT VPN voice traffic, study the impact of the sample sizes and the deep learning methods on identification performance. To verify the performance of the proposed approach, we collect 10 types of typical OTT VPN voice traffic from the experimental 5G network. Extensive experimental results prove the effectiveness of the proposed approach in encrypted 5G OTT VPN voice traffic classification.

Apps installed on smartphones may reveal users' privacy, which is often under malicious attacks. Most privacy attacks are based on network layer traffic. However, the encryption used in the cellular network makes it difficult for a passive adversary to obtain the traffic. In this paper, we leverage the data link layer metadata, such as PDCP packet size, distribution, and interarrival time, to create Apps fingerprints and then conduct a non-intrusive smartphone Apps privacy attack. We test three different smartphones on a 4G LTE laboratory network. On twenty popular Apps selected from the AppStore and Huawei AppGallary, we achieve an Fl-score ranging from 91.32% to 99.49%. Also, we investigate the effect of classification algorithms, time windows, monitoring duration and smartphone brands on Apps fingerprinting attack. Furthermore, we evaluate the performance of the attack using only downlink traffic, which is consistent with the actual attack scenario. Because the data link layer specifications of 4G LTE and 5G are similar, the method of Apps fingerprinting attack can be extended to the latest 5G networks.