Yujun Zhang

With the rapid proliferation of edge computing, mobile edges interconnected via networks are increasingly ex posed to a wide range of attacks. Considering privacy concerns, federated edge anomaly detection, collaboratively training a global detection model across multiple edge clients without centralizing sensitive data, is a promising paradigm to detecting such attacks. However, mobile edge environment is inherently dynamic, with frequent evolving in both network traffic and participating clients. Most existing methods are pre-defined for specific attacks and fixed client settings, which significantly limits their adaptability in dynamic mobile edges. To address these issues, we define the federated adaptability from both data-level and client-level perspectives for the first time, and derive the decisive factors for improving adaptability, i.e., high smoothness and low gradient similarity. Driven by this theoretical foundation, we propose DIFF, a federated edge anomaly detection framework based on dual fuzzing strategies. DIFF incorporates two novel designs for improving adaptability: (i) a classless fuzzing strategy to improve data-level adaptability by fuzzing the boundary between normal and abnormal samples, thereby encouraging model smoothness and enhancing adaptability to unseen traffic and emerging attacks; and (ii) a direction fuzzing strategy to enhance client-level adaptability by perturbing the optimization directions of local models, enabling the aggregated global model to adapt effectively to new clients. Experiments on public datasets demonstrate the superior adapt ability of DIFF compared to the state-of-art methods.

With the emergence of next-generation video applications and increasing spatial resolutions, delivering high-quality video is still limited by network bandwidth. Adaptive bitrate (ABR) can select the appropriate bitrate for video streaming based on bandwidth, which can mitigate rebuffering caused by insufficient bandwidth. In comparison to Constant Bitrate (CBR), Variable Bitrate’s (VBR) encoding scheme can achieve the same quality with less bandwidth consumption and is gradually being widely used in ABR streaming. However, the quality of the video is still degraded due to a poor network. Recent research utilizes Super-resolution (SR) in ABR streaming to construct neural Video Quality Enhancement (VQE) systems, thereby improving the quality of video segments downloaded due to insufficient bandwidth. However, SR cannot participate in the downsampling encoding process of videos, which results in the effectiveness of existing SR-based VQE systems being inherently limited due to unavoidable information loss during downsampling encoding. Concurrently, SR’s high computational cost restricts neural VQE systems’ deployment on clients without GPUs. In contrast to the unidirectional workflow of SR, Rescaling can be integrated into the downsampling encoding process of videos, allowing favorable information to be retained for VQE. To implement high-quality real-time VQE for ABR streaming of VBR-encoded videos on CPUs, we propose RePC, a novel neural VQE system framework for optimizing existing neural VQE systems based on Rescaling (Re) for the first time, and Patch Content-awareness (PC). In detail, RePC uses Rescaling instead of SR to achieve better VQE by participating in the video downsampling. We also propose a Video Single-Image Rescaling model, VSIR, to indicate the effectiveness of RePC in quality enhancement. To speed up VQE, RePC designs a PC algorithm to mix interpolation and neural computation based on the practical upsampling ability. Our evaluation results demonstrate quality gains of 0.55–2.96 dB in PSNR and 1.79–3.18 in VMAF with fewer parameters, a speed-up of 15×–286× well up to real-time requirements on CPUs, and Quality of Experience (QoE) improvements of 16.58–26.65 are also achieved in an ABR system under various networking conditions.

Segment routing (SR) is an emerging architecture that can benefit traffic engineering (TE). Nowadays, TE in SR networks (SR-TE) is often solved as an optimization problem to optimize network performance such as link utilization. As network size grows rapidly, implementing SR-TE suffers from scalability issues, including long computation time, high control overhead and expensive deployment cost. In this paper, we propose Smart Node Selection (SNS), a scalable SR-TE method with learning-based node selection (NS). NS is a recently proposed technique for reducing computation time of SR-TE. It first selects a subset of nodes as candidate intermediate nodes to route traffic, then builds linear programming (LP) models that can be solved efficiently. However, existing NS methods use simple heuristics and consider only network topology, which may lead to unsatisfying network performance. To address this problem, we for the first time formulates NS as a reinforcement learning task, which learns a selection policy to achieve better trade-offs between TE performance and computation time, considering both topology and traffic. Besides, we extend NS with additional selection policies and a customized training algorithm, making it a unified framework for scalable SR-TE, which reduces not only computation time, but also control overhead and deployment cost. Performance evaluations on various real-world topologies and traffic matrices show that SNS significantly reduces computation time and control overhead of existing LP models while offering good network performance, and can also be used in partially deployed SR networks to reduce deployment cost.

As a promising architectural design for future Internet, Named Data Networking (NDN) relies on data names, instead of destination IP addresses, to deliver data. NDN supports data authenticity and integrity by making public key signatures mandatory on data content and data names. This handles the primary security concern in NDN, but is still vulnerable to new DDoS attacks, including Cache Pollution attacks and Interest Flooding attacks, which degrade NDN transmission significantly, by violating the crucial components of NDN routers. To defend against DDoS attacks in NDN, the most effective way is to persistently detect the malicious traffic and then throttle them. Except for the usual concern of the accuracy and efficiency in attack detection, since these attacks themselves have already imposed a huge burden on victims, to avoid exhausting the remaining resources on the victims for detection purpose, a lightweight detection solution is highly desired. We study DDoS attacks and propose a persistent detection solution based on an observed malicious traffic pattern, which leverages a novel sketch to monitor the malicious traffic in a timely and lightweight way. Additionally, our analysis and experiments demonstrate that, with fixed low resource consumption, the proposed solution can persistently detect DDoS attacks in NDN.