Yakun Zhu, Yutong Huang, Shengqian Qin, Zhongzhen Huang, Shaoting Zhang, Xiaofan Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Abstract The Onion Router (Tor), a cornerstone of online privacy, is increasingly exploited for malicious purposes, creating an urgent need to distinguish its traffic from benign web streams. In response to detection efforts, Tor deploys pluggable transports like Obfs4, which obfuscates traffic using randomized padding and artificial timing delays to evade traditional analysis. Although these transformations obscure surface-level patterns, Obfs4 traffic retains subtle yet exploitable statistical divergences from standard web behavior. Critically, existing detection approaches often fail to achieve the throughput necessary for practical, real-time deployment in high-speed network environments, presenting a significant performance gap. To bridge this performance gap, we introduce CTS-OD : A C ascad-ed T wo- S tage framework for high-throughput O bfs4 D etection, engineered for both speed and precision. Our methodology is founded on a compact yet potent set of expressive features selected from early-flow packet data. The first stage implements a label-guided clustering strategy to generate class centroids; these centroids are then indexed using Facebook AI Similarity Search for exceptionally rapid similarity matching, allowing for the immediate classification of the majority of samples. The second stage employs a highly optimized tree-based fallback classifier, which is specifically trained to resolve complex instances that remain ambiguous after the initial assessment. This synergistic architecture achieves an F1-score surpassing 99% and an inference throughput exceeding million samples per second, satisfying the demanding requirements of high-precision, ultra-high-speed inference and making it eminently suitable for real-world deployment.

Traditional collision avoidance methods for unmanned underwater vehicles (UUVs) are highly dependent on the measurement accuracy of sensors and environmental maps and perform poorly in unknown dynamic environments. To improve the generalization ability of the UUV collision avoidance method when there is no prior knowledge of the underwater environment and sonar observation is uncertain, an improved soft actor-critic (SAC) method is proposed in this work. The improved SAC network architecture includes an encoder layer, a dynamic feature layer, and a merge layer. The long short-term memory (LSTM) network is integrated into the dynamic feature layer to capture historical state spaces and action spaces, and improve the collision policy’s memory capability. In this method, the state space is composed of forward-looking sonar perception, UUV motion state, and relative target position. In addition, to enhance the training efficiency and stability of the improved SAC, we design the continuous state space, the action space, and a composite reward function. We built a simulation platform and demonstrated that the improved SAC could effectively complete autonomous collision avoidance in environments with static and dynamic obstacles under external disturbances. The improved method enables the UUV to reach the target faster and reduces the path length and decision-making time of collision avoidance.

This paper proposes a novel deep reinforcement learning (DRL) based method for unmanned underwater vehicle (UUV) path-following and collision avoidance. An end-to-end architecture integrating guidance law, gradient-based condition for action policy smoothing method and soft actor-critic (Grad-CAPS-SAC) is proposed. Specifically, the Serret-Frenet frame and the guidance law generate the tracking error and the desired attitude, respectively. Since the oscillation of the control signal may cause wear and energy consumption of the physical system, an adaptive gradient-based action policy regularization method is proposed to smooth the control policy. In addition, the DRL controller controls the rudder and elevator, and Proportional-integral-derivative (PID) controller is adopted to maintain the desired cruise speed. Furthermore, a composite reward function with constraints is designed to avoid useless exploration and improve training efficiency. Simulation results show that Grad-CAPS-SAC has better path-following and collision avoidance performance than the mainstream DRL methods in different waypoint and obstacle configurations.

Memory disaggregation has attracted great attention recently because of its benefits in efficient memory utilization and ease of management. So far, memory disaggregation research has all taken one of two approaches: building/emulating memory nodes using regular servers or building them using raw memory devices with no processing power. The former incurs higher monetary cost and faces tail latency and scalability limitations, while the latter introduces performance, security, and management problems.

Recently, there is an emerging trend to move towards a disaggregated hardware architecture that breaks monolithic servers into independent hardware components that are connected to a fast, scalable network [1, 2]. The disaggregated architecture offers several benefits over traditional monolithic server model, including better resource utilization, ease of hardware deployment, and support for heterogeneity. Our vision of the future disaggregated architecture is that each component will have its own controller to manage its hardware and can communicate with other components through a fast network.