Low-altitude economy applications, such as unmanned logistics, urban air mobility, and emergency response, are expanding rapidly. Existing terrestrial networks face challenges in coverage, security, and energy efficiency. To address these issues, we propose PLS-MARL, a framework combining physical-layer security with multi-agent reinforcement learning for secure and efficient cross-domain communications. We design a Satellite-enhanced Low-altitude Economy and Terrestrial Network architecture, integrating satellites, unmanned aerial vehicles, and ground nodes to provide seamless 3D coverage and dynamic resource collaboration. Within this architecture, an adaptive physical-layer security mechanism is developed, leveraging artificial noise, stochastic geometry modeling, and cooperative interference to protect against eavesdropping, jamming, and spoofing attacks. To align incentives among heterogeneous nodes, we introduce a Stackelberg game and contract-based scheme that encourages voluntary cooperation while balancing security and energy consumption. At the core, a model-free two-Scale Stackelberg multi-agent reinforcement learning algorithm enables hierarchical decision-making across fast and slow timescales. It dynamically optimizes trajectories, power allocation, and noise injection without prior knowledge of attack models or channel states. Simulation results show that PLS-MARL achieves secrecy outage below 0.15, reduces energy consumption by 32%, and maintains secure availability with only 3% degradation under strong jamming. It also demonstrates stable convergence and robust performance under dynamic and intensive adversarial conditions.
To overcome the growing performance gap between processor speeds and memory access times, commonly known as the memory wall, this paper introduces a solution that uses approximate computing. The proposed method predicts the values of memory load instructions using a pre-trained machine learning model. These predicted values are used in place of actual memory loads, allowing the processor to bypass latency-critical memory operations. The proposed Machine Learning-based Load Value Approximator (ML-LVA) is implemented as a lightweight accelerator within a out-of-order RISC-V processor. Experimental results using image and audio processing workloads show that the ML-LVA achieves up to 73% speedup in memory operations and up to 34.8% in the overall the application. The hardware overhead of deploying the ML-LVA remains modest, with area and power increases of only 5.09% and 0.7%, respectively.
Secure multi-party computation (SMC) has received significant attention recently. As one of the most common operators, sorting is essential in data mining and analysis. To achieve secure multi-party sorting (SMS), many SMS schemes have been proposed. However, most existing schemes neither support deduplication sorting nor achieve efficient communication and computation. To balance security, functionality, and efficiency, we propose an efficient secure multi-party multi-data deduplication sorting scheme, called SMDS. Specifically, the concrete functionality realized by SMDS is deduplicated ranking under set semantics: for each user value, it returns its position in the deduplicated union of all users' datasets, where equal values share the same position. Based on symmetric homomorphic encryption (SHE), we design a multi-party SHE scheme, called Multi-SHE, and combine it with a polynomial-based encoding (PBE) method to support privacy-preserving deduplicated counting and ranking computation over a bounded ordered domain in a non-colluding two-server model with an ideal communication pattern. We further formalize the security of SMDS under a static semi-honest model with explicit leakage and prove that it securely realizes the target functionality assuming the semantic security of Multi-SHE. Moreover, we conduct extensive comparison experiments to evaluate the performance of SMDS, and the results show that it achieves favorable efficiency. Specifically, in a dataset with 18,000 data values for 18 participants, where each participant holds a dataset of size 1,000, the computational cost of SMDS is 40.85 s, which is only 54.39% and 56.34% of that of two related schemes, respectively.
This paper deals with volume diagnosis of memory cores and illustrates how to quickly and accurately classify failure bitmaps dumped from volume production with high confidence. The method predominantly uses Convolutional Neural Networks (CNNs) and permits the quick and accurate classification of the failure shapes into predefined labels. Additionally, the proposed method enables pinpointing new, unclassified failing shapes. Experimental results focus on Non-Volatile Memories (NVMs), which reach sizes up to several megabytes per bank embedded in the System-on-Chip (SoC). The proposed method enables a massive analysis of production lots, being able to crunch thousands of chips and hundreds of thousands of embedded NVM cores per day. The classification accuracy reaches up to 98.64 with an average compute time of less than one hour for a set of around ten thousand Megabit-sized failing NVMs. The method has also proven effective in guiding the identification of outlier shapes.
Surveillance is an essential task in the field of computer vision. Environmental factors like lack of proper light and degradation due to water in underwater videos cause image degradation and, hence, loss of information. State-of-the-art methods can identify objects but do not provide complete information about their shape and structure. This paper proposes Multiple-Aggregator Graph learning for Foreground Segmentation (MAG-FS), a latent graph refactoring framework designed to preserve object boundaries and structural details during foreground segmentation. The spatial features are retrieved using a Convolutional Neural Network (CNN). To reduce the information loss caused by convolution and pooling operations, MAG-FS projects the latent representation into graph space without relying on a predefined adjacency prior. The information in the nodes is combined using task-adaptive statistical aggregators, including mean, variance, and standard deviation, together with degree-aware scalers to enhance important foreground features and suppress background noise. The graph is transformed back into a latent vector during the refactoring process, which is subsequently fed to an image generation module to generate the binary foreground segmentation mask. The proposed technique is evaluated using two underwater datasets, Fish4Knowledge and underwater change detection, and a terrestrial database, CDnet, covering diverse challenges such as dynamic background, bad weather, shadows, turbulence, night videos, and low-frame-rate sequences. The proposed method achieves competitive or superior performance compared with the evaluated state-of-the-art methods. The proposed model achieves 0.99 F-measure on the Fish4Knowledge dataset and 0.98 F-measure on the CDnet dataset. The average inference time of the proposed scheme is 0.029 sec, making the system viable for real-time applications.
