ABSTRACT As urban sound environments grow more complex, environmental monitoring increasingly demands not just sound‐level measurement but also source attribution. Current source recognition techniques can identify which sources are present in a mixture, but cannot measure each source's energy contribution, making accountability impossible. Sound source separation overcomes this by isolating individual sources, yet most deep learning methods assume a fixed number of sources, making them unsuitable for environmental attribution. This paper introduces a blind source separation framework based on spectrogram time–frequency localisation, inspired by the human auditory cortex (organ of Corti). The framework dynamically determines the number of separable sources via encoder‐conditioned queries and isolates each source using its time–frequency coordinates. Two key challenges—mirroring and masking—are analysed. Extensive experiments show that the proposed paradigm significantly outperforms state‐of‐the‐art methods in separation accuracy, computational efficiency and memory usage.
ABSTRACT Accurate segmentation of medical images, particularly along ambiguous lesion boundaries, plays a vital role in clinical diagnosis and treatment planning. Conventional segmentation approaches often struggle to recover fine‐grained contours due to spatial detail loss caused by convolutional downsampling and the inherent sparsity of edge features, leading to incomplete or imprecise delineations. To address these limitations, we propose a feedback‐driven uncertainty‐aware framework, termed UFU‐Net, which iteratively quantifies prediction uncertainty and progressively refines features in low‐confidence regions. Specifically, (i) an uncertainty module interprets the weak decoder's outputs to produce heatmaps that explicitly direct attention to ambiguous areas; (ii) a bidirectional feature‐fusion module aggregates multi‐scale uncertainty cues during decoding to strengthen edge representations; and (iii) a local Manhattan self‐attention mechanism builds pixel‐level context in contour regions, enabling sub‐pixel detail reconstruction. Experiments on ISIC2018 , BUSI and Glas show consistent improvements over state‐of‐the‐art methods, and qualitative results confirm accurate localisation and refinement of irregular lesion contours.
ABSTRACT Delegation learning is an implementation approach within privacy‐preserving machine learning (PPML). It refers to that the delegator outsources the training task to servers with strong computing power due to its own limited computing resources, and the servers return the correct training results and get the service revenue. However, it is difficult to balance the security and usability of delegation learning, which brings great technical challenges to the design of security mechanisms for delegation learning. In this paper, we propose a privacy‐preserving, efficient and publicly verifiable scheme for rational delegation learning (PPEPV‐RDL). First, we demonstrate that in the absence of an effective incentive mechanism, rational participants in delegation learning inevitably fall into a prisoner's dilemma, which leads to a collapse of protocol execution where no participant performs correctly. Second, we design an incentive mechanism and construct a game‐theoretic model for PPEPV‐RDL. We prove the existence of a unique sequential equilibrium, which ensures that rational participants have no incentive to deviate from honest behaviour, thereby achieving incentive compatibility. Finally, we combine the universal composability (UC) theory to construct the ideal function of rational delegation learning and prove the security of PPEPV‐RDL. Simulation experiments show that PPEPV‐RDL not only achieves public verifiability but also maintains data privacy while improving computational efficiency and reducing computational overhead.
ABSTRACT In recent years, research on triangle counting in large graphs has primarily focused on estimating the number of triangles in static graphs. Most streaming graph sampling algorithms are only capable of estimating triangle counts in static graphs and cannot handle dynamic graphs. Meanwhile, existing dynamic streaming graph sampling algorithms suffer from low estimation accuracy. To address these issues, we propose VHADS (Vertex Hashing Assignment‐based Distributed Sampling), a distributed sampling algorithm designed for estimating the number of triangles in large‐scale dynamic graph streams within a cluster environment. As a distributed streaming algorithm, VHADS effectively organises the edge data processed by worker nodes through a vertex‐hash allocation strategy, ensuring that each triangle is sampled by only one worker node in the cluster. This significantly increases the probability of triangle sampling and effectively reduces estimation errors. Compared to state‐of‐the‐art streaming sampling algorithms, VHADS offers the following advantages: (1) With the same sample size, VHADS achieves lower estimation errors in a shorter runtime, reducing the average error in global triangle counting by 72.05% and the average error in local triangle counting by 69.31%; (2) VHADS provides an unbiased estimation of triangle counts in dynamic streaming graphs, and rigorous theoretical analysis demonstrates that it achieves a higher triangle sampling probability.
ABSTRACT Multi‐objective optimisation is crucial in engineering and real‐world decision‐making. Although traditional decomposition‐based approaches such as MOEA/D yield reasonable results, they depend heavily on expert knowledge. Existing MetaBBO approaches partially alleviate this reliance, however, they offer limited configurability and lack real‐world validation. To address these limitations, we propose EMO‐CRL, a highly configurable Evolutionary Multi‐objective Optimisation framework through Collaborative Reinforcement Learning. We first construct a comprehensive configuration space encompassing multiple advanced MOEA/D mechanisms, effectively creating a unified structure for exploring diverse algorithmic variants. Recognising the inherent complexity of the high‐dimensional configuration space, EMO‐CRL then innovatively decomposes it into six sub‐spaces to enable efficient exploration. This decomposition is coupled with the formulation of dynamic algorithm configuration as a multi‐agent Markov Decision Process, featuring comprehensive state representation and finer‐grained action space. EMO‐CRL employs Value‐Decomposition Networks to collaboratively adapt agents across diverse scenarios, allowing effective knowledge sharing and improved decision‐making. Extensive evaluations on WFG and DTLZ benchmarks show that EMO‐CRL outperforms several state‐of‐the‐art MOEA/D variants. Furthermore, its application to UAV path planning highlights its practical effectiveness.
