Collaborative Research: CNS: Medium: Scalable Learning from Distributed Data for Wireless Network Management
Synopsis
As we transition to 5G and beyond, we observe both the emergence of new applications (e.g., Internet of Things, virtual and augmented reality) as well as an increased attack surface exploitable by malicious behaviors (e.g., as evidenced by the Mirai botnet). Future wireless networks will need better control and management; e.g., allocating spectrum, and in identifying and isolating distributed attacks.
In this context, data-driven and machine-learning (ML) based approaches offer tremendous promise to simplify and enhance these management tasks. However, this promise is especially hard to realize in wireless networks, as the data is decentralized (e.g., available at geographically apart base stations). In spite of advances, the backhaul network is fundamentally constrained. Canonical approaches (e.g., sampling or sending raw data) induce fundamental undesirable tradeoffs between scalability and fidelity.
Our focus in this proposal is on developing decentralized telemetry and analytics capabilities that provide scalable and high-fidelity inputs for wireless network management, rather than the control capabilities. The driving principle in our work is to look for algorithmic opportunities minimize the amount of backhaul traffic to send the appropriate telemetry information. The fundamental challenge lies in supporting rich, diverse tasks that often need network-wide visibility, while still keeping the overhead low. To this end, we will develop the algorithmic, machine learning, and network systems foundations to systematically balance these tradeoffs.
The proposed work will be accomplished through the following three convergent tasks and can facilitate a variety of applications such as detecting new patterns, and sophisticated attacks along three tasks:
- Task 1: Automatic generation of sketches/summaries: Towards capturing the telemetry information drawn from local analytics we will develop sketching/summarization approaches. In addition, we propose sketches/summaries that can be used in holistic, federated analytics, and approaches to ensure that heterogeneous sketches can be coalesced for global analytics.
- Task 2: Online federated learning for dynamic updates of local models: To support analytics that need global models or global views, we leverage federated learning for local analytics based telemetry, using a model derived based on views from multiple vantage points. We especially target dynamically evolving local federated models during online operations and deriving a single model for multiple analytics tasks (e.g., determining global AR or IoT traffic distributions. detecting distributed but known DDoS attacks etc.)
- Task 3: Dynamic Adaptation to backhaul bandwidth availability: Finally, we propose ML-based methods to dynamically allocate backhaul bandwidth to the different local base stations, for the transferring the telemetry information, and protocols for locally fine-tuning of the sketches/summaries to adhere to the assigned bandwidth budgets.
Personnel
- PIs:
- Graduate Students:
- Abdelrahman Fahim (UCR)
- Taghreed Alanazi (UCR)
- Xiangyu Chang (UCR)
- Yajie Zhou (BU)
- Antonis Manousis (CMU)
Collaborators
Broader Impacts
- Multiple Ph.D students are either fully or partially supported by the project.
- The use of decentralized data to learn patterns in next generation wireless networks will significantly improve the ability to manage such networks and deliver significantly enahnced QoE to users.
- A new AI/ML summer camp has been organized with the Redlands Unified School District
- We co-organized the data science challenge (DSC) in collaboration with Livermore National Laboratory for students with little or no prior experience in ML.
Publications
- [Milcom] Clustered Federated Learning for Massive MIMO Power Allocation, Y. Yan, X. Chang, A. Roy Chowdhury, S. V. Krishnamurthy, A. Swami and B. Guler, IEEE Milcom 2024.
- [ICALP] Lower Bounds for Pseudo-Deterministic Counting in a Stream, Vladimir Braverman, Robert Krauthgamer, Aditya Krishnan, Shay Sapir. Appeared in ICALP 2023: 30:1-30:14
- [Sensors] Streaming Quantiles Algorithms with Small Space and Update Time Nikita Ivkin, Edo Liberty, Kevin Lang, Zohar Karnin, Vladimir Braverman. Appeared in Sensors. 2022; 22(24):9612
- [SOSR] Flow-Level Loss Detection with Δ-Sketches,Shir Landau Feibish, Zaoxing Liu, Nikita Ivkin, Xiaoqi Chen, Vladimir Braverman, Jennifer Rexford Appeared in SOSR 2022
- [NSDI] Sketchovsky: Enabling Ensembles of Sketches on Programmable Switches, Hun Namkung, Zaoxing Liu, Daehyeok Kim, Vyas Sekar and Peter Steenkiste in USENIX NSDI23
- [VLDB] Enabling Efficient and General Subpopulation Analytics In Multidimensional Data Streams, Antonis Manousis, Zhuo Cheng, Ran Ben Basat, Zaoxing Liu, Vyas Sekar in VLDB22
- [NSDI] SketchLib: Enabling Efficient Sketch-based Monitoring on Programmable Switches
Hun Namkung, Zaoxing Liu, Daehyeok Kim, Vyas Sekar, Peter Steenkiste in USENIX NSDI22
- [IMC] Precise Error Estimation for Sketch-based Flow Measurement
Peiqing Chen, Yuhan Wu, Tong Yang, Junchen Jiang, Zaoxing Liu in ACM/SIGCOMM IMC21
Software