Dynamic Adaptive DNN Surgery for Inference Acceleration on the Edge
The DOI matches its Crossref record; 22 of the 23 references it cites were found in a registry or an index.
Abstract
Recent advances in deep neural networks (DNNs) have substantially improved the accuracy and speed of a variety of intelligent applications. Nevertheless, one obstacle is that DNN inference imposes heavy computation burden to end devices, but offloading inference tasks to the cloud causes transmission of a large volume of data. Motivated by the fact that the data size of some intermediate DNN layers is significantly smaller than that of raw input data, we design the DNN surgery, which allows partitioned DNN processed at both the edge and cloud while limiting the data transmission. The challenge is twofold: (1) Network dynamics substantially influence the performance of DNN partition, and (2) State-of-the-art DNNs are characterized by a directed acyclic graph (DAG) rather than a chain so that partition is greatly complicated. In order to solve the issues, we design a Dynamic Adaptive DNN Surgery (DADS) scheme, which optimally partitions the DNN under different network condition. Under the lightly loaded condition, DNN Surgery Light (DSL) is developed, which minimizes the overall delay to process one frame. The minimization problem is equivalent to a min-cut problem so that a globally optimal solution is derived. In the heavily loaded condition, DNN Surgery Heavy (DSH) is developed, with the objective to maximize throughput. However, the problem is NP-hard so that DSH resorts an approximation method to achieve an approximation ratio of 3. Real-world prototype based on self-driving car video dataset is implemented, showing that compared with executing entire the DNN on the edge and cloud, DADS can improve latency up to 6.45 and 8.08 times respectively, and improve throughput up to 8.31 and 14.01 times respectively.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474
Going deeper with convolutions
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., & Rabinovich, A. (2015) · 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2015.7298594 · Cited by 48,562
YOLO9000: Better, Faster, Stronger
Redmon, J., & Farhadi, A. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.690 · Cited by 19,892 · cited 4×
“For example, the input data size of tiny YOLOv2 [4] is 0 .”
BackgroundInfluentialInception-v4, Inception-ResNet and the Impact of Residual Connections on Learning
Szegedy, C., Ioffe, S., Vanhoucke, V., & Alemi, A. (2017) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v31i1.11231 · Cited by 4,636 · cited 1×
“6 shows the DAG of the pure inception v4 network [10] in Fig.”
BackgroundAn experimental comparison of min-cut/max- flow algorithms for energy minimization in vision
Boykov, Y., & Kolmogorov, V. (2004) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2004.60 · Cited by 4,577
MAUI: making smartphones last longer with code offload
Cuervo, E., Balasubramanian, A., Cho, D.-K., Wolman, A., Saroiu, S., Chandra, R., & Bahl, P. (2010) · Proceedings of the 8th international conference on Mobile systems, applications, and services · doi:10.1145/1814433.1814441 · Cited by 2,281
Dynamic adaptive streaming over HTTP -: standards and design principles
Stockhammer, T. (2011) · Proceedings of the second annual ACM conference on Multimedia systems · doi:10.1145/1943552.1943572 · Cited by 1,221 · cited 1×
“This can be realized by a method similar to HTTP DASH [11].”
MethodDeep neural networks for small footprint text-dependent speaker verification
Variani, E., Lei, X., McDermott, E., Moreno, I. L., & Gonzalez-Dominguez, J. (2014) · 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp.2014.6854363 · Cited by 1,035
Distributed Deep Neural Networks Over the Cloud, the Edge and End Devices
Teerapittayanon, S., McDanel, B., & Kung, H. T. (2017) · 2017 IEEE 37th International Conference on Distributed Computing Systems (ICDCS) · doi:10.1109/icdcs.2017.226 · Cited by 843
(2011) · Cited by 713
Cited by
Items 201–300 of 493
ResMap: Exploiting Sparse Residual Feature Map for Accelerating Cross-Edge Video Analytics
Chen, N., Zhang, S., Zhang, S., Yan, Y., Chen, Y., & Lu, S. (2023) · IEEE Conference on Computer Communications (INFOCOM) · doi:10.1109/infocom53939.2023.10228990 · Cited by 12
Cost Efficient Offloading Strategy for DNN-Based Applications in Edge-Cloud Environment
Huang, Y., Lin, B., Zheng, Y., Hu, J., Mo, Y., & Chen, X. (2019) · IEEE International Symposium on Parallel and Distributed Processing with Applications (ISPA) · doi:10.1109/ispa-bdcloud-sustaincom-socialcom48970.2019.00056 · Cited by 12
“Hu [20] et al designed a Dynamic Adaptive DNN Surgery(DADS) allow DNNs to be processed at the edge and cloud while limiting data transmission.”
BackgroundCheng, Z., Xia, X., Wang, H., Liwang, M., Chen, N., Fan, X., & Wang, X. M. (2025) · IEEE Transactions on Services Computing · doi:10.1109/tsc.2025.3607117 · Cited by 11
“For instance, foundational works such as [19], [20] have established methodologies for model partitioning and inference optimization, while recent systematic surveys [21], [22], [23] offer comprehensive overviews of foundational EI research.”
Exploiting the Computational Path Diversity with In-network Computing for MEC
Guo, X., Dong, F., Shen, D., Huang, Z., Ni, Z., Jiang, Y., & Yin, D. (2022) · IEEE International Conference on Sensing, Communication, and Networking (SECON) · doi:10.1109/secon55815.2022.9918601 · Cited by 11
“[4], [6], [18] split a DNN into two parts.”
BackgroundA Resource-Efficient Feature Extraction Framework for Image Processing in IoT Devices
Ding, C., Li, Y., Lu, Z., Wang, S., & Guo, S. (2022) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2022.3218402 · Cited by 11
“There is also much work on training or inferring models in a cloud-device collaboration fashion, such as [29], [30], [31], [32], [33], [34], [35], [36].”
BackgroundDNN Real-Time Collaborative Inference Acceleration with Mobile Edge Computing
Yang, R., Li, Y., He, H. X., & Zhang, W. (2022) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn55064.2022.9892582 · Cited by 11
“Model partitioning and collaborative computing are current hotspots in edge intelligence research [12, 13, 15, 18, 23–27].”
BackgroundMethodInfluentialVelasco-Montero, D., Goossens, B., Fernández‐Berni, J., Rodriguez-Vazquez, A., & Philips, W. R. (2023) · IEEE Access · doi:10.1109/access.2023.3264828 · Cited by 11
“Strategies based on multiple devices have also been proposed, including hybrid cloud-edge solutions [7], [8], [9], [10], [11], [12], [13] and collaborative systems in the framework of the Internet of Things [14], [15], [16], [17], [18].”
MethodPang, B., Liu, S., Wang, H., Guo, B., Wang, Y., Wang, H., Sheng, Z., Wang, Z., & Yu, Z. (2023) · ACM Transactions on Sensor Networks · doi:10.1145/3630098 · Cited by 11
“To this end, the DNN partition technique [16, 21, 47] becomes an ideal solution to deploy DNNs on heterogeneous mobile and edge clusters.”
BackgroundMethodLearning the Optimal Path and DNN Partition for Collaborative Edge Inference
Huang, Y., Zhang, L., & Xu, J. (2025) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2025.3602966 · Cited by 11
“Prior research on collaborative edge inference has explored DNN partitioning optimization in both offline settings [5], [6], [7], [8], [18], [19] and online settings [9], [10], [20].”
Doubly Residual Neural Decoder: Towards Low-Complexity High-Performance Channel Decoding
Liao, S., Deng, C., Yin, M., & Yuan, B. (2021) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v35i10.17040 · Cited by 11
DeepSave: saving DNN inference during handovers on the edge
Ju, W., Yuan, D., Bao, W., Ge, L., & Zhou, B. B. (2019) · ACM/IEEE Symposium on Edge Computing (SEC) · doi:10.1145/3318216.3363301 · Cited by 11
“DNN Surgery [10] is a DNN partition scheme that dynamically decides the optimal partition cut.”
BackgroundMethodInfluentialSelf-citationLoADPart: Load-Aware Dynamic Partition of Deep Neural Networks for Edge Offloading
Liu, H., Zheng, W., Li, L., & Guo, M. (2022) · IEEE International Conference on Distributed Computing Systems (ICDCS) · doi:10.1109/icdcs54860.2022.00053 · Cited by 11
“DADS[14] proposes a partition decision algorithm that can deal with DNNs with DAG architectures.”
BackgroundReconciling Security and Communication Efficiency in Federated LearningPreprint
Prasad, K., Ghosh, S., Cormode, G., Mironov, I., Yousefpour, A., & Stock, P. (2022) · arXiv (Cornell University) · doi:10.48550/arxiv.2207.12779 · Cited by 11
“Besides gradient compression, other methods such as [52, 26] reduce the communication cost by partitioning the model such that each client learns a portion of it, while [24] proposes training small models and periodically distilling them to a larger central model.”
MethodIntra-DP: A High Performance Collaborative Inference System for Mobile Edge ComputingPreprint
Sun, Z., Guan, X., Lin, Z., Fang, Z., Cai, X., Chen, Z., Liu, F., Cui, H., Xiong, J., Ni, W., et al. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2507.05829 · Cited by 11
“Leveraging the insight from layer partitioning methods [30], whereby certain intermediate DNN layers yield significantly smaller outputs than raw input data, we impose the following constraint to curtail this overhead.”
Self-citationLiao, Z., Zhang, X., He, S., & Tang, Q. (2023) · Journal of Network and Computer Applications · doi:10.1016/j.jnca.2023.103720 · Cited by 10
On the impact of black-box deployment strategies for edge AI on latency and model performance
Singh, J., Fallahzadeh, E., Adams, B., & Hassan, A. E. (2026) · Empirical Software Engineering · doi:10.1007/s10664-025-10714-1 · Cited by 10
DNNSplit: Latency and Cost-Efficient Split Point Identification for Multi-Tier DNN Partitioning
Kayal, P., & Leon‐Garcia, A. (2024) · IEEE Access · doi:10.1109/access.2024.3409057 · Cited by 10
“These approaches can be broadly classified into two-tier [7], [8], [9], [10], [12] or three-tier [8], [11], [13] joint inference acceleration techniques.”
BackgroundMethodInfluentialSmart Split: Leveraging TinyML and Split Computing for Efficient Edge AI
Bove, F., & Bedogni, L. (2024) · IEEE/ACM Symposium on Edge Computing (SEC) · doi:10.1109/sec62691.2024.00052 · Cited by 10
“A dynamic adaptive network for efficient edge intelligence [15] can adjust its architecture on the fly.”
EdgeBoost: Confidence boosting for resource constrained inference via selective offloading
Said, N., & Landsiedel, O. (2025) · Computer Networks · doi:10.1016/j.comnet.2025.111437 · Cited by 10
Yang, Y.-T. C., & Wei, H. (2021) · Asia-Pacific Network Operations and Management Symposium (APNOMS) · doi:10.23919/apnoms52696.2021.9562657 · Cited by 10
“…DL layer to multiple devices due to current network condition, [5] introduced more design knobs that further improve device-edge collaboration, [6] designed a scheme that can partition both directed acyclic graph (DAG) and chain DL model, and [7] provided a complete analysis for representative…”
BackgroundA Low-code Development Framework for Cloud-native Edge Systems
Zhang, W., Zhang, Y., Fan, H., Gao, Y. C., & Dong, W. (2023) · ACM Transactions on Internet Technology · doi:10.1145/3563215 · Cited by 10
“In order to decide which modules to include, we conduct an investigation of more than twenty edge-computing-related articles [5, 8, 14, 25, 27, 29, 42, 44–47, 54, 55, 63, 69–73, 76–78, 80] and find that the main functionalities of edge systems for IoT applications include: (1) connector for both…”
MethodHuang, Z., Dong, F., Shen, D., Wang, H., Guo, X., & Fu, S. (2022) · Proceedings of the International Conference on Parallel Processing · doi:10.1145/3545008.3545071 · Cited by 10
“The DNN is treated as a directed acyclic graph or a chain and partitioned into two parts [12] [17], where one is executed locally and the other is transmitted to a shared edge server for further processing to utilize the resources in the edge.”
MethodEnergy-Efficient Cooperative Inference Via Adaptive Deep Neural Network Splitting at the Edge
Labriji, I., Merluzzi, M., Airod, F. E., & Strinati, E. C. (2023) · International Conference on Communications · doi:10.1109/icc45041.2023.10279444 · Cited by 10
“In [7], the authors propose a strategy to partition the DNN under different network conditions to minimize the overall processing delay.”
DNN Inference Acceleration with Partitioning and Early Exiting in Edge Computing
Li, C., Xu, H., Xu, Y., Wang, Z., & Huang, L. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-85928-2_37 · Cited by 10
RTCoInfer: Real-Time Collaborative CNN Inference for Stream Analytics on Ubiquitous Images
Zhang, Z., Yang, S., Zhao, C., Ren, X., Yu, H., Han, Q., & Guo, S. (2023) · IEEE Journal on Selected Areas in Communications · doi:10.1109/jsac.2023.3242730 · Cited by 10
“Here, predominating solutions can be mainly categorized as two types: progressive inference with multi-models [10], [11], [12], and distributed inference with model partition [13], [16], [22].”
BackgroundEfficient Distributed Retrieval-Augmented Generation for Enhancing Language Model PerformancePreprint
Liu, S., Zheng, Z., Huang, X., Wu, F., Chen, G., & Wu, J. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2504.11197 · Cited by 10
“Tosimultaneouslyachieve privacy preservation and low latency in mobile computing while benefiting from the robust computational power of cloud, numerous studies [6, 19, 28, 32, 61] have investigated device-cloud collaborative inference for conventional neural networks.”
Robust DNN Partitioning and Resource Allocation Under Uncertain Inference Time
Nan, Z., Han, Y., Zhou, S., & Niu, Z. (2025) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2025.3619509 · Cited by 9
“On the Raspberry Pi platform, the inference time of GoogleNet is about 0.8 seconds [5], while for tiny YOLOv2, it is up to 1.8 seconds [6].”
InfluentialGemini: a Real-time Video Analytics System with Dual Computing Resource Control
Lu, R., Hu, C., Wang, D., & Zhang, J. (2022) · IEEE/ACM Symposium on Edge Computing (SEC) · doi:10.1109/sec54971.2022.00020 · Cited by 9
“Edge-cloud video analytics systems were developed, e.g., DNN Surgery [3], EdgeBox [40], where the workloads are partitioned between the edge device and the cloud.”
BackgroundSelf-citationAn Edge-Side Real-Time Video Analytics System With Dual Computing Resource Control
Hu, C., Lu, R., Sang, Q., Liang, H., Wang, D., Cheng, D., Zhang, J., Li, Q., & Peng, J. (2023) · IEEE Transactions on Computers · doi:10.1109/tc.2023.3301136 · Cited by 9
“There are orthogonal research directions where video analytics is conducted on pre-stored videos [1], or the real-time videos are sent to the cloud for cloud or edge-cloud analytics [2].”
BackgroundMethodSelf-citationOnline Management for Edge-Cloud Collaborative Continuous Learning: A Two-Timescale Approach
Lin, S., Zhang, X., Li, Y., Joe‐Wong, C., Duan, J., Yu, D., Wu, Y. F., & Chen, X. (2024) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2024.3451715 · Cited by 9
“…a distributed computing hierarchies, where the edge and terminal devices use the shallow part of the neural network for fast and local reasoning; [22] designed a scheme to optimally partition the DNN under different network conditions; and [17] consider sharing parameters of different layers of…”
BackgroundAutomated Deep Neural Network Inference Partitioning for Distributed Embedded Systems
Kreß, F., Annabi, E. M. E., Hotfilter, T., Hoefer, J., Harbaum, T., & Becker, J. (2024) · IEEE Computer Society Annual Symposium on VLSI · doi:10.1109/isvlsi61997.2024.00019 · Cited by 9
“Consequently, there has also been research on layer-wise DNN inference partitioning as shown in Table I. Hu et al. [5] proposed an automated dynamic scheduling approach which takes network conditions into account.”
MethodAccelerating DNN Inference by Edge-Cloud Collaboration
Chen, J., Qi, Q., Wang, J., Sun, H., & Liao, J. (2021) · IEEE International Performance, Computing, and Communications Conference · doi:10.1109/ipccc51483.2021.9679434 · Cited by 9
Toward Lifelong Unseen Task Processing With a Lightweight Unlabeled Data Schema for AIoT
Tu, T., He, Z., Zheng, Z., Zheng, Z., Jiang, J., Gong, Y., Hu, C., & Cheng, D. (2024) · IEEE Internet of Things Journal · doi:10.1109/jiot.2024.3396282 · Cited by 9
Self-citationrunData: Re-distributing Data via Piggybacking for Geo-distributed Data Analytics over Edges
Jin, Y., Qian, Z., Guo, S., Zhang, S., Jiao, L., & Lu, S. (2021) · IEEE Transactions on Parallel and Distributed Systems · doi:10.1109/tpds.2021.3086274 · Cited by 9
“DADS [38] optimally partitioned the DNN under different network conditions.”
Privacy-preserving Security Inference Towards Cloud-Edge Collaborative Using Differential PrivacyPreprint
Wang, Y., Chen, X., & Wang, Q. (2022) · arXiv (Cornell University) · doi:10.48550/arxiv.2212.06428 · Cited by 9
“Further, Hu et al [8] propose the model partitioning method for directed acyclic graph (DAG) topological DNN models, but it is inefficient for extensive DNN model partitioning.”
BackgroundMethodAccelerating Deep Neural Network Tasks Through Edge-Device Adaptive Inference
Zhang, X., Teng, Y., Wang, N., Sun, B., & Hu, G. (2023) · IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC) · doi:10.1109/pimrc56721.2023.10293996 · Cited by 8
“DADS [6] considers the DAG topologies of DNN models, which transforms the minimization delay problem into a min-cut problem.”
BackgroundMethodHu, Y., Xu, X., Qi, L., Zhou, X., & Xia, X. (2024) · IEEE Transactions on Artificial Intelligence · doi:10.1109/tai.2024.3366880 · Cited by 8
“In terms of model partition, Hu et al. [15] proposed an adaptive DNN partition approach based on Max-Flow/Min-Cut algorithm [16] to minimize the end-to-end latency of DNN inference.”
Distributed LSTM-based Slice Resource Allocation for Beyond 5G Networks
Ehsanian, A., & Spyropoulos, T. (2024) · 2012 International Conference on Computing, Networking and Communications (ICNC) · doi:10.1109/icnc59896.2024.10556119 · Cited by 8
Efficient and privacy-preserving deep inference towards cloud–edge collaborative
Wang, Y., Zhong, G., Duan, Y., Cheng, Y., Yin, M., & Yang, R. (2025) · Applied Soft Computing · doi:10.1016/j.asoc.2025.113381 · Cited by 8
Bandwidth-Aware Adaptive Codec for DNN Inference Offloading in IoT
Xie, X., Zhou, N., Zhu, W., & Liu, J. (2022) · Lecture notes in computer science · doi:10.1007/978-3-031-19839-7_6 · Cited by 8
“Another line of work [20], [16] splits the DNN model at a “bottleneck” to form a lightweight head model deployed at the mobile devices and a heavier tail model at the server.”
Cost-Efficient Sharing Algorithms for DNN Model Serving in Mobile Edge Networks
Dai, H., Wu, J., Wang, Y., Yen, J., Zhang, Y., & Xu, C. (2023) · IEEE Transactions on Services Computing · doi:10.1109/tsc.2023.3247049 · Cited by 8
“With MEC, the DNN workloads can be (partially) pushed to the edge of the cloud network, which in turn could effectively mitigate the strict requirements on the communication, and consequently, fully unleash the potentials of edge computing for cost reduction [10], [11].”
BackgroundAccuracy-Based Task Offloading and Resource Allocation for Edge Intelligence in IoT
Fan, W., Chen, Z., Su, Y., Wu, F., Tang, B., & Liu, Y. (2021) · IEEE Wireless Communications Letters · doi:10.1109/lwc.2021.3128911 · Cited by 8
OCDST: Offloading Chained DNNs for Streaming Tasks
Gao, G., Wu, L., Shao, Z., Yang, Y., & Lin, Z. (2021) · Globecom · doi:10.1109/globecom46510.2021.9685971 · Cited by 8
A Customized Reinforcement Learning based Binary Offloading in Edge Cloud
Li, Y., Chen, L., Zeng, D., & Gu, L. (2020) · IEEE International Conference on Parallel and Distributed Systems (ICPADS) · doi:10.1109/icpads51040.2020.00055 · Cited by 8
AdaMM: Adaptive Object Movement and Motion Tracking in Hierarchical Edge Computing System
Kim, J., Lee, J., & Kim, T. (2021) · Sensors · doi:10.3390/s21124089 · Cited by 8
“The dynamic adaptive DNN surgery (DADS) scheme proposed by Hu et al. [18] aims to optimally partition a DNN network based on periodic monitored network conditions.”
MethodSynergy of Edge Intelligence and Blockchain: A Comprehensive SurveyPreprint
Wang, X., Ren, X., Qiu, C., Xiong, Z., Yao, H., & Leung, V. C. M. (2021) · doi:10.36227/techrxiv.14724360.v1 · Cited by 8
Scheduling LLM Inference with Uncertainty-Aware Output Length PredictionsPreprint
Zheng, H., Zhang, Y., Fu, F., Zhou, X., Luo, H., Zhu, H., Zhu, Y., Wang, H., Yan, X., & Jiang, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.00499 · Cited by 8
“This requires schedulers that can accurately identify such requests to ensure Quality of Service (QoS) and system stability (Hu et al., 2019).”
A Scene-Aware Model Adaptation Scheme for Cross-Scene Online Inference on Mobile Devices
Li, Y., Zhu, H., Deng, Z., Cheng, Y., Zheng, Z., Zhang, L., Chang, S., & Guo, M. (2025) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2025.3574766 · Cited by 7
“Another branch is to divide DNNs and perform collaborative inference on both edge devices and the cloud [13], [14], [15], or to transmit compressed sensory data to the cloud for data recovery and model inference [16], [17].”
Task-Aware Collaborative Inference and Fine-Grained DNN Partitioning in MEC Networks
Zhang, G., Zhang, Q., Feng, L., Zhou, F., Donta, P. K., & Dustdar, S. (2026) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2025.3650680 · Cited by 7
“On one hand, research on partitioning complex-structured DNNs frequently relies on mathematical approaches such as graph min-cut, which are computationally expensive [10], [11], [12].”
Energy-Aware Satellite-Ground Co-Inference via Layer-Wise Processing Schedule Optimization
Chen, Y., Zhang, Q., Xing, R., Li, Y., Ma, X., Yu, C., Zhang, Y., Zhou, A., & Wang, S. (2024) · Asia-Pacific Symposium on Internetware · doi:10.1145/3671016.3674811 · Cited by 7
DNN Partitioning and Assignment for Distributed Inference in SGX Empowered Edge Cloud
Li, Y., Zeng, D., Gut, L., Guo, S., & Zomaya, A. Y. (2024) · IEEE International Conference on Distributed Computing Systems (ICDCS) · doi:10.1109/icdcs60910.2024.00065 · Cited by 7
“Much recent effort has been devoted to optimizing the distributed DNN inference performance in edge cloud, objective at inference acceleration [7], request throughput promotion [8] and cost minimization [5].”
BackgroundMethodInfluentialA Collaborative Cloud-Edge Approach for Robust Edge Workload Forecasting
Li, Y., Zhao, P., Ma, X., Yuan, H., Fu, Z., Xu, M., & Wang, S. (2024) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2024.3502683 · Cited by 7
“However, the current DNN partition is mainly applied to model inference [39], [40] and considered individually for each edge site, which requires a dedicated offline profiling phase to measure the network condition, the processing ability of the edge site, and the computing capacity of cloud server…”
CrowdHMTware: A Cross-Level Co-Adaptation Middleware for Context-Aware Mobile DL Deployment
Liu, S., Guo, B., Luo, S., Wang, Y., Luo, H., Fang, C., Yuan, X., Ma, K., Li, Y., & Yu, Z. (2025) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2025.3549399 · Cited by 8
NEUKONFIG: Reducing Edge Service Downtime When Repartitioning DNNs
Majeed, A. A., Kilpatrick, P., Spence, I. T. A., & Varghese, B. (2021) · IEEE International Conference on Cloud Engineering · doi:10.1109/ic2e52221.2021.00026 · Cited by 7
Fast and accurate edge resource scaling for 5G/6G networks with distributed deep neural networks
Γιαννακάς, Θ., Spyropoulos, T., & Smid, O. (2022) · IEEE International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM) · doi:10.1109/wowmom54355.2022.00021 · Cited by 7
“Recent works that have proposed systematic ways of compressing the information of the transmitted (towards the cloud) maps, e.g., by dynamically adjusting the sent features depending on network conditions [18], [29], [30].”
Intelligence Inference on IoT Devices
Zhang, Q., Li, Y., Zhang, D., Murturi, I., Pujol, V. C., Dustdar, S., & Wang, S. (2023) · Springer eBooks · doi:10.1007/978-3-031-50514-0_9 · Cited by 7
Embedded Distributed Inference of Deep Neural Networks: A Systematic Review
Peccia, F. N., & Bringmann, O. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2405.03360 · Cited by 7
“By throughput, we refer to the number of inputs that the system can process per second, which is typically calculated as stated in [53], where the throughput of the entire system is simply the inverse of the maximum latency across the groups of layers executed on each device.”
BackgroundRodríguez, I., Campos, C., & Fdez‐Riverola, F. (2023) · Neurocomputing · doi:10.1016/j.neucom.2023.126835 · Cited by 7
Learning to Listen... On-Device: Present and future perspectives of on-device ASR
Vipperla, R., Ishtiaq, S. S., Li, R., Bhattacharya, S., Leontiadis, I., & Lane, N. D. (2020) · GetMobile Mobile Computing and Communications · doi:10.1145/3400713.3400715 · Cited by 7
Synergistic and Efficient Edge-Host Communication for Energy Harvesting Wireless Sensor NetworksPreprint
Mishra, C. S., Sampson, J., Kandmeir, M. T., Narayanan, V., & Das, C. R. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2408.14379 · Cited by 7
“…communication, power-requirement and quality of service (QoS), have pursued three major approaches: inference effort partitioning optimizations [22, 30, 31, 50], mitigation of energy provisioning limitations [24, 40, 43, 47, 56], and minimizing communication overheads [32, 33, 36, 37, 45].”
BackgroundTian, J., Li, X., & Qin, X. (2024) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn60899.2024.10651115 · Cited by 6
Energy-Efficient, Delay-Constrained Edge Computing of a Network of DNNs
Ghasemi, M., Heidari, S., Kim, Y. G., Wu, C.-J., & Vrudhula, S. B. K. (2024) · IEEE Transactions on Computers · doi:10.1109/tc.2024.3500368 · Cited by 6
“The authors in [21] propose a solution to partition the computation of a single DNN modeled as a directed acyclic graph (DAG) between an edge device and a cloud server.”
LVMScissor: Split and Schedule Large Vision Model Inference on Mobile Edges via Salp Swarm Algorithm
Liu, Y., Lu, R., & Wang, D. (2025) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2025.3550519 · Cited by 6
Self-citationOptimizing DNNs Model Partitioning for Enhanced Performance on Edge Devices
Maruf, M. A., & Azim, A. (2023) · Canadian AI · doi:10.21428/594757db.acb1ea67 · Cited by 6
“Several recent studies [8–12] have explored the possibility of DNN partitioning on different computing devices (such as edge devices and CPUs/GPUs) to improve training and inference times.”
BackgroundCan 5G mmWave Enable Edge-Assisted Real-Time Object Detection for Augmented Reality?
Ghoshal, M., Kong, Z. J., Xu, Q., Lu, Z., Aggarwal, S., Khan, I., Meng, J., Li, Y., Hu, Y., & Koutsonikolas, D. (2023) · Proceedings - International Symposium on Modeling, Analysis, and Simulation of Computer and Telecommunication Systems · doi:10.1109/mascots59514.2023.10387645 · Cited by 6
“…the arrival of 5G networks, the research community had already intensively studied edge-assisted object detection ( e.g., [8], [3], [2], [9], [10], [11]) and developed a host of sophisticated “application-level” optimizations in trying to make edge-assisted AR achieve acceptable performance…”
BackgroundAnole: Adapting Diverse Compressed Models for Cross-Scene Prediction on Mobile Devices
Li, Y., Zhu, H., Deng, Z., Cheng, Y., Zhang, L., Chang, S., & Guo, M. (2024) · IEEE International Conference on Distributed Computing Systems (ICDCS) · doi:10.1109/icdcs60910.2024.00063 · Cited by 6
SecoInfer: Secure DNN End-Edge Collaborative Inference Framework Optimizing Privacy and Latency
Yao, Y., Hou, J., Wu, G., Cheng, Y., Yuan, M., Luo, P., Wang, Z., & Li, X. (2024) · ACM Transactions on Sensor Networks · doi:10.1145/3694972 · Cited by 6
“This enhances the inference efficiency of DNNs in sensor networks [15, 17, 25, 56].”
InfluentialAccelerating on-device DNN inference during service outage through scheduling early exit
Wang, Z., Bao, W., Yuan, D., Ge, L., Tran, N. H. K., & Zomaya, A. Y. (2020) · Computer Communications · doi:10.1016/j.comcom.2020.08.005 · Cited by 6
Self-citationPark, J., Kim, J., & Ko, J. H. (2021) · Sensors · doi:10.3390/s21030896 · Cited by 6
“As IoT edge devices enhance their AI processing capabilities, the offloading mechanism has evolved to partial offloading, called ‘Collaborative Intelligence’ [9–12].”
BackgroundJoint Model, Task Partitioning and Privacy Preserving Adaptation for Edge DNN Inference
Jiang, J., Li, H., & Wang, L. (2022) · IEEE Wireless Communications and Networking Conference · doi:10.1109/wcnc51071.2022.9771620 · Cited by 6
“…to the states of the IoT device, such as the transmission rate and CPU/GPU utilization; then we can compute the DNN layers ahead of the partitioning point on the IoT device, and transfer relatively little intermediate results (IRs) to the edge server for computing the residual DNN layers [4].”
MethodEfficient Distributed Inference of Deep Neural Networks via Restructuring and Pruning
Abdi, A., Rashidi, S., Fekri, F., & Krishna, T. (2023) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v37i6.25815 · Cited by 6
“For example, DNN-Surgery (Hu et al. 2019) and similar works consider each layer of DNN as a whole, neither tries to break the layer into sub-layers nor restructures the model.”
BackgroundUnlocking Efficiency: Understanding End-to-End Performance in Distributed Analytics Pipelines
Souza, A., Ng, N., Abdelzaher, T., Towsley, D., & Shenoy, P. J. (2023) · IEEE Military Communications Conference (MILCOM) · doi:10.1109/milcom58377.2023.10356378 · Cited by 6
“Various research work has proposed a layer-wise DNN partitioning approach to distribute the processing between sensors and edges [16], [17].”
BackgroundScission: Context-aware and Performance-driven Edge-based Distributed Deep Neural Networks.Preprint
Lockhart, L., Harvey, P., Imai, P., Willis, P., & Varghese, B. (2020) · arXiv (Cornell University) · Cited by 6
Edge-PRUNE: Flexible Distributed Deep Learning InferencePreprint
Boutellier, J. J., Tan, B., & Nurmi, J. (2022) · arXiv (Cornell University) · doi:10.48550/arxiv.2204.12947 · Cited by 6
“DADS [20] proposed the directed acyclic graph (DAG) based edge-cloud DNN inference model (ECDI) that can be used to optimize the edge-cloud partitioning of more complex than chain-like DNN structures.”
Synergy of Edge Intelligence and Blockchain: A Comprehensive SurveyPreprint
Wang, X., Ren, X., Qiu, C., Xiong, Z., Yao, H., & Leung, V. C. M. (2021) · doi:10.36227/techrxiv.14724360 · Cited by 6
A Resource-Efficient Multiple Recognition Services Framework for IoT Devices
Ding, C., Liu, Z., Zhou, A., Yu, J., Li, Y., & Wang, S. (2024) · IEEE Transactions on Services Computing · doi:10.1109/tsc.2024.3512949 · Cited by 5
“A large number of excellent cloud-based training neural network models have sprung up [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24].”
Dynamic Collaborative Inference for Multi-Type DNN Tasks in Space-Ground Networks: A DRL Approach
Tian, S., Wang, R., Hao, J., Wu, Q., & Niyato, D. T. (2025) · IEEE Transactions on Cognitive Communications and Networking · doi:10.1109/tccn.2025.3616114 · Cited by 5
“…channel conditions and stringent bandwidth constraints, direct transmission of raw image data using traditional satellite-ground collaborative computing schemes would result in end-to-end latency exceeding task deadlines [12], while energy consumption risks surpassing onboard power capacities.”
Split Federated Learning for Resource-Constrained Edge Computing Networks
Wang, X., Song, S., Zhang, Z., Hou, X., Li, Z., Xing, T., & Zhang, X.-P. (2025) · IEEE Transactions on Consumer Electronics · doi:10.1109/tce.2025.3626460 · Cited by 5
“Under dynamic wireless conditions, [20] further explored optimal DNN split locations for delay minimization.”
Partitioning multi-layer edge network for neural network collaborative computing
Li, Q., Zhou, M., Ren, T.-F., Jiang, C.-B., & Chen, Y. (2023) · EURASIP Journal on Wireless Communications and Networking · doi:10.1186/s13638-023-02284-x · Cited by 5
Zheng, Y., Cui, L., Tso, F. P., Li, Z., & Jia, W. (2024) · Computer Networks · doi:10.1016/j.comnet.2024.110607 · Cited by 5
Lin, J., Li, M., Zhang, S. Q., & Leon‐Garcia, A. (2024) · Proceedings of the International Conference on Parallel Processing · doi:10.1145/3673038.3673154 · Cited by 5
“[5 ] goes a step further to design a Dynamic Adaptive DNN Surgery Scheme that generates an optimal partition of a DNN based on dynamic network conditions.”
MethodReview of Collaborative Inference in Edge Intelligence: Emphasis on DNN Partition
Liu, N., Dai, F., Chai, X., Liu, G., Wu, X., & Huang, B. (2024) · IEEE Intl Conf on Dependable, Autonomic and Secure Computing, Intl Conf on Pervasive Intelligence and Computing, Intl Conf on Cloud and Big Data Computing, Intl Conf on Cyber Science and Technology Congress (DASC/PiCom/CBDCom/CyberSciTech) · doi:10.1109/cyberscitech64112.2024.00014 · Cited by 5
“For example, in [4], authors designed a DADS scheme to optimize partitions the DNN under different network condition.”
FastTuner: Fast Resolution and Model Tuning for Multi-Object Tracking in Edge Video Analytics
Xu, R., Nalaie, K., & Zheng, R. (2025) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2025.3526573 · Cited by 5
“More fine-grained partitioning strategies are explored in [43], [44], [45], [46], [47], [48], [49], [50], [51], [52], [53], [54], [55], where neural network layers are split, and features of the intermediate layer are offloaded to another executor for the remaining inference.”
Inference Routing Over Multi-Hop Edge Networks
Xu, C., Liu, Y., & Yang, J. (2025) · IEEE Transactions on Cognitive Communications and Networking · doi:10.1109/tccn.2025.3613490 · Cited by 5
“The authors in [8] and [17] model DNN as a graph and optimize the optimal split points.”
MJOA-MU: End-to-edge collaborative computation for DNN inference based on model uploading
Yang, H., Sun, S., Liu, M., Zhang, Q., & Wang, Y. (2023) · Computer Networks · doi:10.1016/j.comnet.2023.109801 · Cited by 5
Xu, W., Yin, Y., Chen, N., & Tu, H. (2023) · International Journal of Software Engineering and Knowledge Engineering · doi:10.1142/s0218194023410085 · Cited by 5
RobustEdge: Low Power Adversarial Detection for Cloud-Edge Systems
Moitra, A., Bhattacharjee, A., Kim, Y., & Panda, P. (2024) · IEEE Transactions on Emerging Topics in Computational Intelligence · doi:10.1109/tetci.2024.3360316 · Cited by 5
“The Gaussian Noise (GN) attack is created by adding gaussian noise with standard deviation of strenght ϵ to the input.”
BackgroundLatency Minimization for Split Federated Learning
Guo, J., Xu, C., Ling, Y., Liu, Y., & Yu, Q. (2023) · IEEE Vehicular Technology Conference · doi:10.1109/vtc2023-fall60731.2023.10333442 · Cited by 5
“Under different channel conditions, optimal DNN splitting is determined for both delay minimization and throughput maximization in [11].”
MethodShi, L., Xu, Z., Shi, Y., Fan, Y., Ding, X., & Sun, Y. (2021) · Lecture notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering · doi:10.1007/978-3-030-67537-0_15 · Cited by 5
State-of-the-art Techniques in Deep Edge IntelligencePreprint
Lodhi, A. H., Akgün, B., & Özkasap, Ö. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2008.00824 · Cited by 5
“In [50], the authors present the Dynamic Adaptive Surgery Scheme (DADS) to segment DNNs under various network conditions by treating inference minimization as a min-cut problem.”
Channel-Aware Throughput Maximization for Cooperative Data Fusion in CAVPreprint
An, H., Fang, Z., Zhang, Y., Hu, S., Chen, X., Xu, G., & Fang, Y. M. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2410.04320 · Cited by 5
“For example, Google’s autonomous vehicle is capable of amassing up to 750 megabytes of sensor data per second [57].”
Yao, Y., Wang, Z., Luo, P., Cheng, Y., Hou, J., & Li, X. (2025) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2025.3607483 · Cited by 4
“Due to the significant disparity in computational power between edge nodes and end devices, existing strategies often assign minimal tasks to the end device, resulting inalargevolumeof intermediatedata—or eventheoriginal data-uploaded [11], [12], [13], [14], [20], [21].”
Chen, Y., Wang, H., Liu, J. P., Jia, X., Zhao, J., & Qiu, H. (2025) · IEEE International Conference on Distributed Computing Systems (ICDCS) · doi:10.1109/icdcs63083.2025.00076 · Cited by 4
Tang, S., Xia, X., Bilal, M., Dou, W., & Xu, X. (2025) · IEEE Transactions on Consumer Electronics · doi:10.1109/tce.2025.3627545 · Cited by 4
“Considering the structure characteristics of CNN models, the inference could be executed in a parallel manner and start at any layer [46], [47] [48].”
Membership inference attacks against split inference via knowledge transfer
Yi, W., Yang, M., Wang, J., Hu, H., Li, Z., Xu, X., He, Y., & Wang, Y. (2025) · Neurocomputing · doi:10.1016/j.neucom.2025.132247 · Cited by 4
Computing Tasks Saving Schemes Through Early Exit in Edge Intelligence-Assisted Systems
Niu, X., Lv, X., Chen, W., Yu, C., & Jin, H. B. (2025) · IEEE Transactions on Computers · doi:10.1109/tc.2025.3533098 · Cited by 4
“The model partitioning partitions the DNN model into two components, which are executed on the end device and the edge server respectively, with the aim of reducing the latency or energy consumption for edge devices in completing computing tasks [11], [23], [24], [25], [26], [27].”
Luo, H., Tian, H., Zhang, P., Zhao, F., Zhou, M. C., & Sun, Y. (2021) · International Conference on Space-Air-Ground Computing (SAGC) · doi:10.1109/sagc52752.2021.00029 · Cited by 4
Multi-exit DNN inference acceleration for intelligent terminal with heterogeneous processors
Zhang, J., Xin, W., Lv, D., Wang, J., Cai, G., & Dong, F. (2023) · Sustainable Computing Informatics and Systems · doi:10.1016/j.suscom.2023.100906 · Cited by 4
ecoBLE: A Low-Computation Energy Consumption Prediction Framework for Bluetooth Low Energy
Schuhmacher, L., Pollin, S., & Sallouha, H. (2023) · Lirias · doi:10.48550/arxiv.2309.16686 · Cited by 4
“This makes it unsuitable for the current trend in IoT, where machine learning models are implemented on computationally-limited edge devices [8].”
BackgroundEfficient Federated Digital Twin Synchronization in Edge-Cloud Collaborative System
Kim, J.-W., Je-Gal, H., & Lee, H.-S. (2023) · IEEE International Conference on Metaverse Computing, Networking and Applications (MetaCom) · doi:10.1109/metacom57706.2023.00134 · Cited by 4
“Also, the transmission time at each partition point is set to increase as the partition point is located towards the front of the DDNN model, since typically a smaller amount of data needs to be transmitted as the input data is processed via the DNN model [15].”
Method
Metadata
- Authors
- Hu, C., Bao, W., Wang, D., & Liu, F.
- Affiliations
- Hong Kong Polytechnic University (HK)
- The University of Sydney (AU)
- Venue
- IEEE INFOCOM 2019 - IEEE Conference on Computer Communications
- Publisher
- IEEE
- Published
- 2019-04
- Pages
- 1423-1431
- OpenAlex
- W2920031528
- Cited by
- 416
- References
- 23
- Citations by year
- 2026: 29 · 2025: 61 · 2024: 65 · 2023: 92 · 2022: 62 · 2021: 72 · 2020: 26 · 2019: 9
- FWCI
- 36.52
- Type
- conference paper
- Language
- en
- Open access
- closed
- Licence
- Locations
- IEEE Conference on Computer Communications (INFOCOM) (publishedVersion)
- Rare & Special e-Zone (The Hong Kong University of Science and Technology) (submittedVersion)
- Topics
- Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Hardware and Architecture › Cloud Computing and Remote Desktop Technologies
- Keywords
- DNN partitioning, cloud-edge collaboration, inference acceleration, dynamic network conditions, directed acyclic graphs, minimum cut, approximation algorithms, throughput maximization, latency reduction, data transmission reduction
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.