Illegal related-party transactions (RPT) are federal felonies that pose a severe threat to the stability and integrity of modern financial systems. The increasing frequency of RPTs forms complex and dynamic networks. Existing temporal graph learning methods tend to treat entities as functionally homogeneous, ignoring the diverse and evolving structural roles of nodes. Role-based embedding methods model global structure by bridging same-role nodes, but their reliance on a unified mechanism for aggregation and evolution means they fail to distinguish the underlying logic of distinct interactions governed by structural roles. The limitations motivate us to develop a customized role-based strategy. It can also adapt to evolving RPT dynamics, thereby forming a continuous regulatory process to combat illegal activities. In this paper, we propose an innovative Role Perceptual Augmented Temporal Graph Network (RPATGN) for proactive RPT detection. We analyze the structural roles of nodes and employ a role-based message passing mechanism that adapts its aggregation strategy based on the roles of interacting nodes. We integrate a variational graph recurrent neural network, enhanced by temporal contextual attention, to explicitly model the dynamics of the roles and the overall network evolution. Extensive experiments on real-world financial datasets demonstrate the effectiveness of our approach for RPT detection. It holds practical significance for fostering robust financial systems and promoting healthy, transparent financial markets.

Deploying multimedia deep neural networks (DNNs) on edge devices reduces retrieval and inference latency, but untrusted hardware in this setting can easily leak model parameters. Existing TEE-based protections have limitations: the partial-weight obfuscation methods are vulnerable due to statistical flaws, while full-weight obfuscation struggles to balance security and efficiency. To address these, we propose a convolution decomposition obfuscation scheme (MirageNet) for protecting edge multimedia DNNs in TEE–GPU heterogeneous environments. This scheme relieves the flaw that cosine similarity remains highly consistent between pre-trained and fine-tuned models. It obfuscates via convolution-kernel element-wise operations, decoy-kernel injection, and channel/kernel permutations, maximizing GPU utilization while minimizing TEE overhead. Experiments show that the MirageNet scheme reduces attack success to a black-box level, lowers runtime overhead by 15% compared to SOTA, and preserves inference accuracy identical to the original model—meeting practical edge multimedia retrieval demands.

The uncontrollable nature of deployment environments and supply chains for edge devices makes the security of deployed deep neural network (DNN) models a key concern. Leveraging the hardware isolation provided by Trusted Execution Environments (TEEs) to protect sensitive layers of the model is considered a practical solution. However, current studies face the following challenges: (1) The dynamic memory management of mainstream deep learning frameworks is incompatible with the static memory allocation mechanism of TEEs, hindering the integration of an efficient intelligent computing ecosystem throughout the model lifecycle. (2) TEEs lack native support for parallel computing, leading to a performance gap between the TEE and the Rich Execution Environment (REE), which increases overall inference latency. To address these issues, this paper proposes a TEE-based model inference scheme integrated with PyTorch for TrustZone-enabled edge devices. The solution adopts a pre-allocated operator management strategy to eliminate the incompatibility between dynamic graph features and the static memory allocation in the TEE. A semantic-segmentationbased parallel optimization is also introduced to reduce the inference latency of sensitive layers within the TEE. We implement a prototype system on Phytium and evaluate it using four well-known DNN models. Experimental results show that, compared to existing TEE-based DNN inference solutions, our design reduces the lines of code (LoC) for model construction scripts by an average of of 82.2%, lowers secure memory overhead by up to 58.5%, and improves inference performance by an average of$5.38 \times$via octa-core parallel optimization.

The approximate nearest neighbor (ANN) search in high-dimensional space offers a wide spectrum of applications across many domains such as database, machine learning, multimedia and computer vision. A variety of ANN search algorithms have been proposed in the literature. In recent years, proximity graph-based approaches have attracted considerable attention from both industry and academic settings due to the superior search performance in terms of speed and accuracy. A recent work utilizes a graphics processing unit (GPU) to accelerate the ANN search on proximity graphs. Though significantly reducing the distance computation time by taking advantage of the massive parallelism of GPUs, the algorithm suffers from the high expenses of data structure operations. In this paper, we propose a novel GPU -accelerated algorithm that designs a novel GPU-friendly search framework on proximity graphs to fully exploit the massively parallel processing power of GPUs at key steps of the search. Also, we propose GPU-accelerated proximity graph construction algorithms which can build high-quality representative proximity graphs with efficient parallel implementations. Extensive experiments on benchmark high-dimensional datasets demonstrate the outstanding performance of our proposed algorithms in both ANN search and proximity graph construction.

Matrix Factorization (MF) has been widely applied in machine learning and data mining. Due to the large computational cost of MF, we aim to improve the efficiency of SGD-based MF computation by utilizing the massive parallel processing power of heterogeneous multiprocessors. The main challenge in parallel SGD algorithms on heterogeneous CPU-GPU systems lies in the strategy to assign tasks. We design a novel strategy to divide the matrix into a set of blocks by considering two aspects. First, we observe that the matrix should be divided nonuniformly, and relatively large blocks should be assigned to GPUs to saturate the computing power of GPUs. In addition to exploiting the characteristics of hardware, the workloads assigned to two types of hardware should be balanced. We design a cost model tailored for our problem to accurately estimate the performance of hardware on different data sizes. Extensive experiments show that our proposed algorithm achieves high efficiency with a high quality of training quality.