论文检索

Hongtao Huang, Xiaojun Chang, Wen Hu 0001, Lina Yao 0001

Recent years have seen the explosion of edge intelligence with powerful Deep Neural Networks (DNNs). One popular scheme is training DNNs on powerful cloud servers and subsequently porting them to mobile devices after being lightweight. Conventional approaches manually specialized DNNs for various edge platforms and retrain them with real-world data. However, as the number of platforms increases, these approaches become labour-intensive and computationally prohibitive. Additionally, real-world data tends to be sparse-label, further increasing the difficulty of lightweight models. In this paper, we propose MatchNAS, a novel scheme for porting DNNs to mobile devices. Specifically, we simultaneously optimise a large network family using both labelled and unlabelled data and then automatically search for tailored networks for different hardware platforms. MatchNAS acts as an intermediary that bridges the gap between cloud-based DNNs and edge-based DNNs.

Qian Chen 0032, Zilong Wang 0001, Jiaqi Hu 0003, Haonan Yan, Jianying Zhou 0001, Xiaodong Lin 0001

Federated learning (FL) is becoming a major driving force behind machine learning as a service, where customers (clients) collaboratively benefit from shared local updates under the orchestration of the service provider (server). Representing clients' current demands and the server's future demand, local model personalization and global model generalization are separately investigated, as the ill-effects of data heterogeneity enforce the community to focus on one over the other. However, these two seemingly competing goals are of equal importance rather than black and white issues, and should be achieved simultaneously. In this paper, we propose the first algorithm to balance personalization and generalization on top of game theory, dubbed PAGE, which reshapes FL as a co-opetition game between clients and the server. To explore the equilibrium, PAGE further formulates the game as Markov decision processes, and leverages the reinforcement learning algorithm, which simplifies the solving complexity. Extensive experiments on four widespread datasets show that PAGE outperforms state-of-the-art FL baselines in terms of global and local prediction accuracy simultaneously, and the accuracy can be improved by up to 35.20% and 39.91%, respectively. In addition, biased variants of PAGE imply promising adaptiveness to demand shifts in practice.

Mingyang Zhou 0001, Gang Liu 0028, Kezhong Lu, Rui Mao 0001, Hao Liao

Federated learning enables collaborative AI training across organizations without compromising data privacy. Decentralized federated learning (DFL) improves this by offering enhanced reliability and security through peer-to-peer (P2P) model sharing. However, DFL faces challenges in terms of slow convergence rate due to complex P2P graphs. To address this issue, we propose an efficient algorithm to accelerate DFL by introducing a limited number of k of edges into the P2P graphs. Specifically, we establish a connection between the convergence rate and the second smallest eigenvalue of the laplacian matrix of the P2P graph. We prove that finding the optimal set of edges to maximize this eigenvalue is an NP-complete problem. Our quantitative analysis shows the positive effect of strategic edge additions on improving this eigenvalue. Based on the analysis, we then propose an efficient algorithm to compute the best set of candidate edges to maximize the second smallest eigenvalue, and consequently the convergence rate is maximized. Our algorithm has a low time complexity of O(krn^2). Experimental results on diverse datasets validate the effectiveness of our proposed algorithms in accelerating DFL convergence.

Yijing Lin, Zhipeng Gao 0001, Hongyang Du 0001, Dusit Niyato, Jiawen Kang 0001, Xiaoyuan Liu 0002

With the development of AI-Generated Content (AIGC), data is becoming increasingly important, while the right of data to be forgotten, which is defined in the General Data Protection Regulation (GDPR) and permits data owners to remove information from AIGC models, is also arising. To protect this right in a distributed manner corresponding to federated learning, federated unlearning is employed to eliminate history model updates and unlearn the global model to mitigate data effects from the targeted clients intending to withdraw from training tasks. To diminish centralization failures, the hierarchical federated framework that is distributed and collaborative can be integrated into the unlearning process, wherein each cluster can support multiple AIGC tasks. However, two issues remain unexplored in current federated unlearning solutions: 1) getting remaining clients, those not withdraw from the task, to join the unlearning process, which demands additional resources and notably has fewer benefits than federated learning, particularly in achieving the original performance via alternative unlearning processes and 2) exploring mechanisms for dynamic unlearning in the selection of remaining clients possessing unbalanced data to avoid starting the unlearning from scratch. We initially consider a two-level incentive and unlearning mechanism to address the aforementioned challenges. At the lower level, we utilize evolutionary game theory to model the dynamic participation process, aiming to attract remaining clients to participate in retraining tasks. At the upper level, we integrate deep reinforcement learning into federated unlearning to dynamically select remaining clients to join the unlearning process to mitigate the bias introduced by the unbalanced data distribution among clients. Experimental results demonstrate that the proposed mechanisms outperform comparative methods, enhancing utilities and improving accuracy.

Daiki Ito, Yuta Takata, Hiroshi Kumagai, Masaki Kamizono

Traditionally, top-level domains (TLDs) are managed by the Internet corporation for assigned names and numbers (ICANN), and the domain names under them are managed by registrars. Against such a centralized management, a blockchain naming service (BNS) has been proposed to manage TLDs on blockchains without authority intervention. BNS users can register TLD strings as non-fungible tokens and manage the TLD root zone. However, such a decentralized management results in the introduction of a new security issue, BNS TLD name collision, wherein the same TLD is registered in several different BNSs. In this study, we investigated BNS TLD name collisions by analyzing TLDs registered on two BNSs: Handshake and Decentraweb. Specifically, we collected TLDs registered in Handshake and Decentraweb and the associated data, and analyzed the data registration status of BNS TLDs and BNS TLD name collisions. The analysis of 11,595,406 Handshake and 11,889 Decentraweb TLDs revealed 6,973 BNS TLD name collisions. In particular, lastname TLDs, which are intended for use as person names, yielded a large number of registered domain names. In addition, the analysis identified 10 name collisions between the BNS and operational ICANN TLDs. Further, the ICANN TLD candidates under review also had name collisions against the BNS TLDs. Consequently, based on the characteristics of these name collisions and discussions in BNS communities, we considered countermeasures against BNS TLD name collisions. For the further development of BNSs, we believe that it is essential to discuss with the existing Internet communities and coexist with the existing Internet.

Zhen Qin 0004, Xueqiang Yan, Mengchu Zhou, Shuiguang Deng

Federated learning (FL) enables the collaborative training of machine learning models without sharing training data. Traditional FL heavily relies on a trusted centralized server. Although decentralized FL eliminates the dependence on a centralized server, it faces such issues as poisoning attacks and data representation leakage due to insufficient restrictions on the behavior of participants, and heavy communication costs in fully decentralized scenarios, i.e., peer-to-peer (P2P) settings. This work proposes a blockchainbased fully decentralized P2P framework for FL, called BlockDFL. It takes blockchain as the foundation, leveraging the proposed voting mechanism and a two-layer scoring mechanism to coordinate FL among participants without mutual trust, while effectively defending against poisoning attacks. Gradient compression is introduced to lower communication cost and to prevent data from being reconstructed from transmitted model updates. The results of extensive experiments conducted on two real-world datasets exhibit that BlockDFL obtains competitive accuracy compared to centralized FL and can defend against poisoning attacks while achieving efficiency and scalability. Especially when the proportion of malicious participants is as high as 40%, BlockDFL can still preserve the accuracy of FL, outperforming existing fully decentralized P2P FL frameworks based on blockchain.

Haozhao Wang, Yabo Jia, Meng Zhang 0045, Qinghao Hu 0004, Hao Ren 0001, Peng Sun 0006, Yonggang Wen 0001, Tianwei Zhang 0004

Sub-model extraction based federated learning has emerged as a popular strategy for training models on resource-constrained devices. However, existing methods treat all clients equally and extract sub-models using predetermined rules, which disregard the statistical heterogeneity across clients and may lead to fierce competition among them. Specifically, this paper identifies that when making predictions, different clients tend to activate different neurons of the entire model related to their respective distributions. If highly activated neurons from some clients with one distribution are incorporated into the sub-model allocated to other clients with different distributions, they will be forced to fit the new distributions, which can hinder their activation over the previous clients and result in a performance reduction. Motivated by this finding, we propose a novel method called FedDSE, which can reduce the conflicts among clients by extracting sub-models based on the data distribution of each client. The core idea of FedDSE is to empower each client to adaptively extract neurons from the entire model based on their activation over the local dataset. We theoretically show that FedDSE can achieve an improved classification score and convergence over general neural networks with the ReLU activation function. Experimental results on various datasets and models show that FedDSE outperforms all state-of-the-art baselines.

Weichen Bi, Yun Ma 0002, Yudong Han 0001, Yifan Chen 0005, Deyu Tian, Jiaqi Du

Graphics rendering on web browsers serves as the foundation for numerous web applications. Compared with the widely employed WebGL, the next-generation web graphics API, WebGPU, demonstrates an enhanced capacity to adapt to modern GPU features, boasting more significant potential. However, our experiment shows that the performance of current graphics rendering frameworks based on WebGPU lags behind those built on WebGL. Such discrepancy primarily arises from an incomplete alignment with WebGPU's distinctive features. The individual rendering of each graphic leads to redundant communication between the CPU and GPU. To enhance the graphics performance on the web, we introduce the FusionRender to harness the power of WebGPU. To mitigate redundant communication, FusionRender assigns a unique signature to each object and employs these signatures for grouping, enabling the consolidation of graphics rendering whenever possible. In simulated experiments involving the rendering of multiple objects, FusionRender improves the rendering performance by 29.3%-122.1% compared with the existing optimal baseline. In real cases with more complex features, performance improvement ranges from 9.4% to 39.7%. Additionally, FusionRender exhibits robust performance enhancement across various devices and browsers.

Jiawei Liu 0007, Yutong Zhai, Gongming Zhao, Hongli Xu 0001, Jin Fang, Zhen Zeng, Ying Zhu

Deep learning has brought about a revolutionary transformation in network applications, particularly in domains like e-commerce and online advertising. Distributed training (DT), as a critical means to expedite model training, has progressively emerged as a key foundational infrastructure for such applications. However, with the rapid advancement of hardware accelerators, the performance bottleneck in DT has shifted from computation to communication. In-network aggregation (INA) solutions have shown promise in alleviating the communication bottleneck. Regrettably, current INA solutions primarily focus on improving efficiency under the traditional parameter server (PS) architecture and do not fully address the communication bottleneck caused by limited PS ingress bandwidth. To bridge this gap, we propose InArt, the first work to introduce INA with routing selection in a multi-PS architecture. InArt employs a multi-PS architecture to split DT tasks among multiple PSs, and selects appropriate routing schemes to fully harness INA capabilities. To accommodate traffic dynamics, InArt adopts a two-phase approach: splitting the training model among multiple parameter servers and selecting routing paths for INA. We propose Lagrange multiplier and randomized rounding algorithms for these phases, respectively. We implement InArt and evaluate its performance through experiments on physical platforms (Tofino switches) and Mininet emulation (P4 Software Switches). Experimental results show that InArt can reduce communication time by 48%\!\sim57\!% compared with state-of-the-art solutions.

Jia Zhang 0010, Haixuan Tong, Enhuan Dong, Xin Qian, Mingwei Xu 0001, Xiaotian Li, Zili Meng

Mobile web services value a quick loading of contents in the first page, which is quantified by the above-the-fold time of the first page (first AFT) and is likely to fall into the slow start phase in congestion control. However, the widely deployed slow start mechanism is "cold start", which manually hardcodes the parameters and is not suitable for the first AFT of heterogeneous mobile web services. We revisit the slow start mechanism and find that it could be optimized with a priori knowledge. However, blindly relying on a priori knowledge is not robust enough to handle the fluctuating mobile networks and unpredictable application traffic. In this paper, we propose WiseStart, a "hot-start-based" slow start mechanism. WiseStart utilizes the priori knowledge to set the initial parameters, continuously probes the new connection to handle the fluctuating network conditions, and carefully adapts to the application-limit scenarios. We implement WiseStart in a popular mobile web service online in production. Comprehensive experiments demonstrate that WiseStart reduces the First AFT by 25.43% and the average RCT at connection establishment by 16.15% compared to the default slow start mechanism and other state-of-the-art baselines.

Zhaoyang Yu 0002, Shenglin Zhang, Mingze Sun, Yingke Li, Yankai Zhao, Xiaolei Hua, Lin Zhu, Xidao Wen, Dan Pei

With the rapid development of cellular networks, wireless base stations (WBSes) have become crucial infrastructure for mobile web systems. To ensure service quality, operators constantly monitor the operation status of WBSes and deploy anomaly detection methods to identify anomalies promptly. After the deployment of anomaly detection methods, operators periodically collect feedback, which holds significant value in improving anomaly detection performance. In real-world industrial environments, the frequency of false negative feedback is usually very low, and the newly generated data's distribution can differ significantly from that of the original training data. Therefore, the feedback-based performance improvement of the previously proposed methods is limited. In this paper, we propose AnoTuner, which incorporates a false negative augmentation mechanism to generate similar false negative feedback cases, effectively compensating for the low feedback frequency. Additionally, we introduce a Two-Stage Active Learning (TSAL) mechanism that minimizes data contamination issues caused by the difference between the distribution of feedback data and that of the training data. Experiments conducted on the real-world data collected from a top-tier global Internet Service Provider (ISP) demonstrate that the performance improvement of AnoTuner after feedback-based fine-tuning is significantly higher than that of the best baseline method.

Zhenge Xu, Qing Li 0006, Wanxin Shi, Yong Jiang 0001, Zhenhui Yuan, Peng Zhang 0104, Gabriel-Miro Muntean

With the rapid popularity of short video applications, a large number of short video transmissions occupy the bandwidth, placing a heavy load on the Internet. Due to the extensive number of short videos and the predominant service for mobile users, traditional approaches (e.g., CDN delivery, edge caching) struggle to achieve the expected performance, leading to a significant number of redundant transmissions. In order to reduce the amount of traffic, we design a Novel Coded Transmission Mechanism (NCTM), which transmits XOR-coded data instead of the original video content. NCTM caches the short videos that users have already watched in user devices, and encodes, multicasts, and decodes XOR-coded files separately at the server, edge nodes, and clients, with the assistance of cached content. This approach enables NCTM to deliver more short video data given the limited bandwidth. Our extensive trace-driven simulations show how NCTM reduces network load by 3.02%-14.75%, cuts peak traffic by 23.01%, and decreases rebuffering events by 43%-85% in comparison to a CDN-supported scheme and a naive edge caching scheme. Additionally, NCTM also increases the user's buffered video duration by 1.21x-13.53x, ensuring improved playback smoothness.

Pengze Li, Mingxuan Song, Mingzhe Xing, Zhen Xiao, Qiuyu Ding, Shengjie Guan, Jieyi Long

Sharding provides an opportunity to overcome the inherent scalability challenges of the blockchain, which is the infrastructure for the next generation of the Web. In a sharding blockchain, the state is partitioned into smaller groups known as "shards." Since the states are placed on different shards, cross-shard transactions are inevitable, which is detrimental to the performance of the sharding blockchain. Existing solutions place states based on heuristic algorithms or redistribute states via graph-partitioning-based methods, which are either less effective or costly. In this paper, we present SPRING, the first deep-reinforcement-learning(DRL)-based sharding framework for state placement. SPRING formulates the state placement as a Markov Decision Process, which considers the cross-shard transaction ratio and workload balancing and employs DRL to learn the effective state placement policy. Experimental results based on real Ethereum transaction data demonstrate the superiority of SPRING compared to other state placement solutions. In particular, it decreases the cross-shard transaction ratio by up to 26.63% and boosts throughput by up to 36.03%, all without unduly sacrificing the workload balance among shards. Moreover, updating the training model and making decisions takes only 0.1s and 0.002s, respectively, which shows the overhead is acceptable.

Haichuan Hu, Fangming Liu, Qiangyu Pei, Yongjie Yuan, Zichen Xu 0001, Lin Wang 0015

Graph Neural Networks (GNNs) have been increasingly adopted for graph analysis in web applications such as social networks. Yet, efficient GNN serving remains a critical challenge due to high workload fluctuations and intricate GNN operations. Serverless computing, thanks to its flexibility and agility, offers on-demand serving of GNN inference requests. Alas, the request-centric serverless model is still too coarse-grained to avoid resource waste. Observing the significant data locality in computation graphs of requests, we propose λGrapher, a serverless system for GNN serving that achieves resource efficiency through graph sharing and fine-grained resource allocation. "Grapher features the following designs: (1) adaptive timeout for request buffering to balance resource efficiency and inference latency, (2) graph-centric scheduling to minimize computation and memory redundancy, and (3) resource-centric function management with fine-grained resource allocation catered to the resource sensitivities of GNN operations and function orchestration optimized to hide communication latency. We implement a prototype of λGrapher based on the representative open-source serverless platform Knative and evaluate it with real-world traces from various web applications. Our results show that λGrapher can achieve an average savings of 61.5% in memory resource and 47.2% in computing resource compared with the state of the arts while ensuring GNN inference latency.

Yuanzheng Tan, Qing Li 0006, Junkun Peng, Zhenhui Yuan, Yong Jiang 0001

Drones connected via the web are increasingly being used for crowd anomaly detection (CAD). Existing solutions, however, face many challenges, such as low accuracy and high latency due to drones' dynamic shooting distances and angles as well as limited computing and networking capabilities. In this paper, we propose Air-CAD, an edge-assisted multi-drone network that uses air-ground cooperation to achieve fast and accurate CAD. Air-CAD consists of two stages: person detection and multi-feature analysis. To improve CAD accuracy, Air-CAD dynamically adjusts the inference of person detection model based on drones' shooting distances and assigns appropriate feature analysis tasks to drones shooting at variable angles. To achieve fast CAD, edge devices connected to drones are deployed to offload assigned feature analysis tasks from drones. Air-CAD schedules the connection between each drone and edge to accelerate processing based on drone's assigned task and the computing/network resources of the edge device. To validate the performance of Air-CAD, we generate a new simulated human stampede dataset captured from various drone-view recordings. We deploy and evaluate Air-CAD in both simulation and real-world testbed. Experimental results show that Air-CAD achieves 95.33% AUROC and real-time inference latency within 0.47 seconds.

Sihao Hu, Tiansheng Huang, Ka-Ho Chow 0001, Wenqi Wei 0001, Yanzhao Wu 0001, Ling Liu 0001

Language models (LMs) have demonstrated superior performance in detecting fraudulent activities on Blockchains. Nonetheless, the sheer volume of Blockchain data results in excessive memory and computational costs when training LMs from scratch, limiting their capabilities to large-scale applications. In this paper, we present ZipZap, a framework tailored to achieve both parameter and computational efficiency when training LMs on large-scale transaction data. First, with the frequency-aware compression, an LM can be compressed down to a mere 7.5% of its initial size with an imperceptible performance dip. This technique correlates the embedding dimension of an address with its occurrence frequency in the dataset, motivated by the observation that embeddings of low-frequency addresses are insufficiently trained and thus negating the need for a uniformly large dimension for knowledge representation. Second, ZipZap accelerates the speed through the asymmetric training paradigm: It performs transaction dropping and cross-layer parameter-sharing to expedite the pre-training process, while revert to the standard training paradigm for fine-tuning to strike a balance between efficiency and efficacy, motivated by the observation that the optimization goals of pre-training and fine-tuning are inconsistent. Evaluations on real-world, large-scale datasets demonstrate that ZipZap delivers notable parameter and computational efficiency improvements for training LMs. Our implementation is available at: https://github.com/git-disl/ZipZap.

Denizhan Kara, Tomoyoshi Kimura, Shengzhong Liu, Jinyang Li 0004, Dongxin Liu, Tianshi Wang 0002, Ruijie Wang 0004, Yizhuo Chen, Yigong Hu, Tarek F. Abdelzaher

This paper presents FreqMAE, a novel self-supervised learning framework that synergizes masked autoencoding (MAE) with physics-informed insights to capture feature patterns in multi-modal IoT sensor data. FreqMAE enhances latent space representation of sensor data, reducing reliance on data labeling and improving accuracy for AI tasks. Differing from data augmentation-based methods like contrastive learning, FreqMAE's approach eliminates the need for handcrafted transformations. Adapting MAE for IoT sensing signals, we present three contributions from frequency domain insights: First, a Temporal-Shifting Transformer (TS-T) encoder that enables temporal interactions while distinguishing different frequency bands; Second, a factorized multi-modal fusion mechanism for leveraging cross-modal correlations and preserving unique modality features; Third, a hierarchically weighted loss function that emphasizes important frequency components and high Signal-to-Noise Ratio (SNR) samples. Comprehensive evaluations on two sensing applications validate FreqMAE's proficiency in reducing labeling needs and enhancing resilience against domain shifts.

Jinliang Yuan, Shangguang Wang, Hongyu Li, Daliang Xu, Yuanchun Li 0003, Mengwei Xu 0001, Xuanzhe Liu

AI is making the Web an even cooler place, but also introduces serious privacy risks due to the extensive user data collection. Federated learning (FL), as a privacy-preserving machine learning paradigm, enables mobile devices to collaboratively learn a shared prediction model while keeping all training data on devices. However, a key obstacle towards practical cross-device FL training is huge energy consumption, especially for lightweight mobile devices. In this work, we perform the first-of-its-kind analysis of improving FL performance through low-precision training with an energy-friendly Digital Signal Processor (DSP) on mobile devices. We first demonstrate that directly integrating the state-of-the-art INT8 (8-bit integer) training algorithm and classic FL protocols will significantly degrade the model accuracy. Moreover, we observe that there are still unavoidable frequent quantization operations on devices that cause extreme load stress on DSP-enabled INT8 training. To address the above challenges, we present Q-FedUpdate, an FL framework that efficiently preserves model accuracy with ultra-low energy consumption. It maintains a global full-precision model and allows the tiny model updates to be continuously accumulated, instead of being erased by the quantization. Furthermore, it introduces pipelining technology to parallel CPU-based quantization and DSP-enabled training, which reduces the floating-point computation overhead of frequent data quantization. Extensive experiments show that Q-FedUpdate can effectively reduce the on-device energy consumption by 21×, and accelerate the FL convergence by 6.1× with only 2% accuracy loss.

Jiaming Huang, Yi Gao 0001, Wei Dong 0001

With the increasing demand for Web of Things (WoT) and edge computing, the efficient utilization of limited computing power on edge devices is becoming a crucial challenge. Traditional neural networks (NNs) as web services rely on deterministic computational resources. However, they may fail to output the results on non-deterministic computing power which could be preempted at any time, degrading the task performance significantly. Multi-exit NNs with multiple branches have been proposed as a solution, but the accuracy of intermediate results may be unsatisfactory. In this paper, we propose MEEdge, a system that automatically transforms classic single-exit models into heterogeneous and dynamic multi-exit models which enables Memory-Elastic inference at the Edge with non-deterministic computing power. To build heterogeneous multi-exit models, MEEdge uses efficient convolutions to form a branch zoo and High Priority First (HPF)-based branch placement method for branch growth. To adapt models to dynamically varying computational resources, we employ a novel on-device scheduler for collaboration. Further, to reduce the memory overhead caused by dynamic branches, we propose neuron-level weight sharing and few-shot knowledge distillation(KD) retraining. Our experimental results show that models generated by MEEdge can achieve up to 27.31% better performance than existing multi-exit NNs.

Lianchen Jia, Chao Zhou 0003, Tianchi Huang, Chaoyang Li 0002, Lifeng Sun

Congestion control has been a fundamental research focus in web transmission for over 30 years. However, with diverse network scenarios like cellular networks and WiFi, traditional models might no longer accurately describe current network conditions -- we empirically observe that the minimum round-trip time (RTTmin) still varies under different network conditions, challenging the assumption of its constancy in traditional models. In this paper, we model it as a normal distribution based on our measurements and propose a novel congestion control algorithm LingBo. LingBo consists of two phases: an offline trained decision model to achieve goals under different RTTmin distributions, and an online perception scheme to detect the current RTTmin distribution. We evaluate LingBo in various network environments and find it consistently performs well in terms of power metric and throughput compared to recent state-of-the-art baselines. Our code is available at https://github.com/thumedia/LingBo.