Threshold fully homomorphic encryption (TFHE) enables secure multi-party computations on encrypted data without revealing sensitive information. Current TFHE constructions primarily rely on the Shamir secret sharing (SSS) scheme or the linear secret sharing scheme (LSSS), each facing challenges such as lack of access structure flexibility and high system complexity. The TreeSSS method addresses these problems by iteratively conducting the SSS, but its efficiency degrades as the number of participants increases, limiting its scalability for largescale system. In this paper, we propose a role-based dynamic scheme for TFHE to improve both the efficiency and flexibility of access control. All valid parties are divided into three roles: L-1 (i.e., data owners), L-2 (i.e., computation executors), and L-3 (i.e., auditors). Our scheme employs a one round SSS with distribute both the TFHE secret key sk and a random mask r to the corresponding parties. The system supports dynamic role transitions without altering sk by refreshing r, thereby maintaining multi-party homomorphic computations. Compared with existing constructions, our method reduces the complexity to O(N log N) and introduces flexible access control.
Hardware-software co-design is crucial in the design of artificial intelligence (AI) inference accelerators. This study supports the development of a variable-parallelism reconfigurable architecture designed to achieve high-efficiency and low-power computation for diverse AI models. We provide software-side design support to achieve high accuracy with limited computational resources. We propose an algorithm that automatically explores hardware configurations by selecting a layer-wise computation flow to achieve high inference accuracy with limited computational resources, which is realized via software-based emulation. The proposed method is employed to reduce weights by 75% through post-training quantization while limiting the Top-1 accuracy degradation to about 3% for image classification on ImageNet-1k with convolutional neural network models such as VGG16. Furthermore, we verify that the proposed exploration method can efficiently identify both the optimal computation flow and minimal hardware requirements required to maintain accuracy.
The multi-coloring method can be formulated as a graph coloring problem, which is known to be NP-hard. In this study, we solve the graph coloring problem using a quadratic unconstrained binary optimization (QUBO) formulation. We perform parameter tuning and propose an iterative color reduction method to further decrease the number of colors. Finally, we evaluate the effectiveness of the resulting parallel ordering as a preprocessing step for iterative methods, compared with conventional approaches.
The demand for high-performance computing (HPC) has been rapidly increasing with the advancement of scientific computing and AI workloads. During peak periods, resource contention often leads to prolonged job waiting times. Cloud Bursting (CB), which elastically extends on-premises resources with on-demand cloud instances, has emerged as a practical solution. However, CB introduces a trade-off between job waiting times and minimizing cloud costs. Effectively balancing these conflicting objectives remains a major challenge. Conventional approaches based on weighted scalarization singleobjective reinforcement learning tend to bias optimization toward one objective, resulting in poor coverage of the Pareto Front (PF). In this study, we apply a multi-objective reinforcement learning framework, Pareto Conditioned Networks (PCN), to HPC job scheduling for preserving the trade-off structure between job waiting time and cloud cost. Through simulation experiments conducted under four different settings that vary the number of jobs and the node size of each resource, PCN consistently outperformed conventional methods. PCN achieved approximately 43% improvement over competing approaches in terms of hypervolume (HV), which measures both the extent and the quality of the obtained PF. Furthermore, PCN maintained lower values of the spacing metric, which evaluates the uniformity of PF distribution. By offering a broader and evenly distributed PF, PCN provides HPC system operators with greater flexibility in adaptively selecting trade-offs according to operational requirements.
An autonomous mobile robot system consists of multiple simple robots that are anonymous, identical, oblivious, and execute the same algorithm. Each robot repeatedly performs a Look-Compute-Movecycle, and clarifying the solvability of tasks under such restrictions is a central issue. The defected view model, introduced by Kim et al., captures faults in the Look operation, where a robot may fail to observe some others. Variants include distance-based, adversarial, and relaxed models. Previous work showed that gathering is possible in the distance-based (4,2)-defected view model under a fully synchronous scheduler, but its solvability under weaker schedulers remained open. This paper proves the impossibility of gathering problem in the distance-based $(4,2)$-defected view model even under a semi-synchronous scheduler, and further establishes the unsolvability of the distinct gathering problem in the same setting.
