Recent foundation models are capable of handling multiple tasks and multiple data modalities with the unified base model structure and several specialized model components. However, efficient training of such multi-task (MT) multi-modal (MM) models poses significant system challenges due to the sophisticated model architecture and the heterogeneous workloads of different tasks and modalities. In this paper, we propose Spindle, a brand new training system tailored for resource-efficient and high-performance training of MT MM models via wavefront scheduling. The key idea of Spindle is to decompose the model execution into waves and address the joint optimization problem sequentially, including both heterogeneity-aware workload parallelization and dependency-driven execution scheduling. We build our system and evaluate it on various MT MM models. Experiments demonstrate the superior performance and efficiency of Spindle, with speedup ratio up to 71% compared to state-of-the-art training systems.

Mission-critical applications require deterministic services for coexisting periodic and bursty flows. While the original cyclic queuing and forwarding (CQF) bounds delay via a dual-queue ping-pong buffering mechanism, its rigid architecture suffers from limited scalability and poor adaptability to traffic bursts. This paper introduces multi-queue CQF—a multi-queue-per-port architecture enabling fine-grained flow isolation—and a co-optimization framework for routing and scheduling with provable guarantees in offline and online regimes. We develop a primal-dual approximation scheme that efficiently determines the maximum schedulable flow set under resource constraints and an online scheduler that ensures high delivery success with only slight capacity augmentation. Numerical evaluations demonstrate that our solutions maintain high schedulability and robustness across diverse optical network configurations.

In the machine learning system, the hybrid model parallelism combining tensor parallelism (TP) and pipeline parallelism (PP) has become the dominant solution for distributed training of Large Language Models~(LLMs) and Multimodal LLMs (MLLMs). However, TP introduces significant collective communication overheads, while PP suffers from synchronization inefficiencies such as pipeline bubbles. Existing works primarily address these challenges from isolated perspectives, focusing either on overlapping TP communication or on flexible PP scheduling to mitigate pipeline bubbles. In this paper, we propose a new synergistic tensor and pipeline parallelism schedule that simultaneously reduces both types of bubbles. Our proposed schedule decouples the forward and backward passes in PP into fine-grained computation units, which are then braided to form a composite computation sequence. This compositional structure enables near-complete elimination of TP-related bubbles. Building upon this structure, we further design the PP schedule to minimize PP bubbles. Experimental results demonstrate that our approach improves training throughput by up to 12\% for LLMs and 16\% for MLLMs compared to existing scheduling methods. Our source code is avaiable at https://github.com/MICLAB-BUPT/STP.

The burgeoning demands of modern applications highlight the imperative of supporting deterministic, ultrareliable, and low-latency services in mobile networks, however, ensuring deterministic delays always faces significant challenges due to the inherent unpredictability of the time-varying channel conditions and the bursty data traffic. In this paper, we investigate online joint user and resource scheduling in unknown dynamic environments, which yields a nested mixed-integer nonlinear programming problem. To deal with such a formidably complex optimization task, we propose a drift-plus-penalty scheme for the higher-level user selection and devise a low-complexity dual method with slight performance degradation for the lower-level resource allocation. Our approach accommodates diverse user deadlines and establishes a tunable tradeoff between power consumption and queue blockage. We provide analysis on convergence and performance bounds, and numerical results demonstrate the effectiveness and the efficiency of our proposal.

Deterministic mobile networks are essential for advanced applications that demand strict quality of service (QoS) assurances under limited resource availability. Though network slicing can optimize average performance metrics to offer best-effort services, it often fails to meet the high-reliability requirements of deterministic communication scenarios. In this paper, we introduce a novel QoS-guaranteed inter-slice radio resource allocation scheme for mobile networks to deliver deterministic services over the long term. First, we develop an analytical martingale-based stochastic network calculus framework, which yields stochastic bounds for transmission delays and queue backlogs across various traffic arrival patterns. These bounds produce robust interval estimations that guide resource allocation decisions, effectively addressing channel variability and long-tail QoS effects. Then, an efficient resource allocation algorithm is proposed to approach the derived performance bounds while ensuring fairness across different radio slices with diverse QoS needs. The framework also incorporates an adaptive traffic predictor, enabling our algorithm to track and respond to network dynamics. Numerical results demonstrate that our proposed scheme achieves a promising trade-off between resource utilization and QoS guarantees.

Deterministic communications systems are critical for time-sensitive applications in the Internet of Things, which demand stringent delay requirements under the conditions of limited available resources. Though network slicing can provide best-effort services by focusing on average performance metrics, it generally cannot address the worse-case issues arising from the deterministic communication scenarios. In this paper, we propose an efficient inter-slice radio resource allocation scheme for mobile networks to provide delay-guaranteed services, where a stochastic network calculus model is introduced to analyze the service delay and its variation. We derive a tight and time-invariant upper bound of the delay violation probability, and develop an efficient resource allocation algorithm to meet the stringent delay requirement of different radio slices. Numerical results demonstrate our proposal achieves a promising trade-off between resource utilization and delay.

With the development of decentralized networks, smart contracts, especially those for ERC tokens, are attracting more and more Dapp users to implement their applications. There are some functions in ERC token contracts that only a specific group of accounts could invoke. Among those functions, some even can influence other accounts or the whole system without prior notice or permission. These functions are referred to as contract backdoors. Once exploited by an attacker, they can cause property losses and harm users’ privacy. In this work, we propose Pied-Piper, a hybrid analysis method that integrates datalog analysis and directed fuzzing to detect backdoor threats in Ethereum ERC token contracts. First, datalog analysis is applied to abstract the data structures and identification rules related to the threats for preliminary static detection. Then, directed fuzzing is applied to eliminate false positives caused by the static analysis. We first evaluated Pied-Piper on 200 smart contracts, which are injected with different types of backdoors. It reported all problems without false positives, and none of the injected problems was missed. Then, we applied Pied-Piper on 13,484 real token contracts deployed on Ethereum. Pied-Piper reported 189 confirmed problems, four of which have been assigned unique CVE ids while others are still in the review process. Each contract takes 8.03 seconds for datalog analysis on average, and the fuzzing engine can eliminate the false positives within one minute.

In order to detect the type of vehicle seat and the missing part of the spring hook, this paper proposes an improved RANSAC-SURF method. First, the image is filtered by a Gauss filter. Second, an improved RANSAC-SURF algorithm is used to detect the types of vehicle seats. Extract the feature points of vehicle seats. The feature points are matched according to the improved RANSAC-SURF algorithm. Third, the image distortion of the vehicle seat is corrected by the method of perspective transformation. Determine whether the seat’s spring hook is missing or not according to the absolute value of the gray difference between the image collected by the camera and the image of the normal installation. The experimental results show that the MSE of the Gauss filter under a 5 [Formula: see text] 5 template is 19.0753, and the PSNR is 35.3261, which is better than that of the mean filter and the median filter. The total matching logarithm of feature points and the number of intersection points are 188 and 18, respectively, in the improved RANSAC-SURF matching algorithm.

A novel laser printing files detection method is proposed in this paper to solve the problem of low efficiency and difficulty in traditional detection. The new method is based on improved scale-invariant feature transform (SIFT) feature and histogram feature. Firstly, analyze the graphical features of different laser printing files. Different files have different printing texture features in valid data area. So segment the valid data area to remove the interference of background. Secondly, extract the histogram feature of the same character in the printing file. Normalize the histogram and then calculate the Bhattacharyya coefficient between the detected file and the original file to determine whether the detected file is right or fake. At the same time, calculate the SIFT features and match the detected file and the original file. To focus on the letter or character region, the SIFT features which are out of contour are deleted. Finally, the results of the two different methods are both used as the result of the identification. When any of the result is fake, the end result will be fake. In the self-built database experiment, in different printing files from different printers, the inkjet areas possess different image features. When scanning different files using 600 dpi, the detect accuracy is higher than 97%. This method was able to meet the reliability requirements of law.

With the constant development of enterprises, the cost of developing new business becomes higher than that in the past, but there is still no better strategy to integrate old businesses together. Comparing integrated tactics such as EAI (enterprise application integration), SOI (service-oriented integration), etc., combining the most popular technology SOA (services oriented architecture), pouring the relevant experience into integration according to the idea proposed by BOA (business oriented architecture) and utilizing different kinds of agents to deal with different businesses. The author will introduce a sandwich model to solve the problem of business integration. And at the end of this paper, the feasibility, advantages, and disadvantages of the model are regarded