Federated Model Reusability at the Edge

Abstract

Our research mainly investigates efficient knowledge management among distributed nodes/users. At the current stage, we specifically study the potential of reusing models with meta-learning at the Edge because of the high possibility of redundant model/data in the Internet of Things era. To select and find suitable models for proper users, we innovate a matching principle based on their tasks and data statistical similarity, called the Borrower-Loaner-Matching process, with a corresponding monitoring mechanism, ensuring the model reusability. Furthermore, a Multitask Learning approach can be applied to reusable models to improve their generalization ability. In addition, we dig into methods to support more complex tasks at distributed nodes the efficient Federated Learning.

Keywords: Model Reuse; Distributed Computing; Distributed Machine Learning; Knowledge Management; Federated Learning; Meta Learning.

Team: Qianyu Long, Christos Anagnostopoulos, Fani Deligianni

Current outcome

Activities

  • Project poster presented in PhD SICSA 2022 conference, Glasgow Caledonian University, 28-29 Jun 2022.

Graphical Abstract

Predictive Learning at the Edge through Distributed Node Selection

Abstract

This research project focuses on distributed learning among nodes in Edge Computing environments. We investigate intelligent edge-nodes selection mechanisms per analytics query involving the collaborative communication patterns among the edge nodes and task off-loading decisions based on uncertainty. Based on the selected nodes, a.k.a. clique, we can build incremental predictive models per query and investigate how this tailored node selection affects models' loss, robustness, complexity, and accuracy. In addition, node selection copes with the availability of nodes' data and requested data for model training and testing influenced by the (local) data diversity in each node.

Key words: Distributed machine learning, incremental learning, Edge computing, collaborative models, task off-loading decision making.

Team:  Tahani Aladwani, Christos Anagnostopoulos, Fani Deligianni, Ibrahim Alghamdi, Kostas Kolomvatsos

Funding: Saudi Ministry of Education

Current Outcome

Activities

  • Paper presentation: Query-driven Edge Node Selection in Distributed Learning Environments [DASC2023], Data-driven Smart Cities (DASC) 2023 in conjuction with the 39th IEEE International Conference on Data Engineering (ICDE 2023), 3rd-7th April 2023, Anaheim, California, USA
  • Project poster ADSAI-2023 presented in Advances in Data Science and AI Conference 2023, University of Manchester, 6 July, 2023.
  • Project poster presented in PhD SICSA 2022 conference, Glasgow Caledonian University, 28-29 Jun 2022.
  • Paper presentation in IEEE The 9th International Conference on Future Internet of Things and Cloud (FiCloud 2022), Italy, 22-24 Aug 2022.
  • Research presented in Global AI Summit, Saudi Arabia, 13-15 September 2022.

 Graphical Abstract

Distributed Data Synopses at the Edge Mesh

Abstract

Internet of Things offers the infrastructure for smooth functioning of autonomous context-aware devices being connected towards the Cloud. Edge Computing (EC) relies between the IoT and Cloud providing significant advantages. One advantage is to perform local data processing (limited latency, bandwidth preservation) with real time communication among IoT devices, while multiple nodes become hosts of the collected data (reported by IoT devices). In this project, we introduce proactive mechanisms for the exchange of data synopses (summaries of extracted knowledge) among EC nodes that are necessary to convey summarized knowledge across EC nodes. The overarching aim is to intelligently decide on when nodes should exchange up-to-date data synopses in light of efficient execution of tasks. We enhance such a decision with stochastic optimization models based on the Theory of Optimal Stopping, thus, obtaining the optimal times/decision to disseminate data synopses to network edge for executing tasks and services. 

Key words: Distributed data synopses, decentralized machine learning, task/service management & off-loading, collaborative predictive analytics.

Team:  Christos Anagnostopoulos, Kostas Kolomvatsos

Funding: EP/R018634/1-UK EPSRC 2018-2023; GNFUV Fire+ EU/H2020

Current Outcome

 Graphical Abstract

 

Resilience at the Edge

Abstract

In distributed Machine Learning (ML) systems that work under the scope of Edge Computing (EC), edge nodes (referred to as nodes hereinafter) are not always reliable and could go offline due to various reasons. We aim to investigate a substitution mechanism that enables peer nodes to serve requests in the place of the failing one by extracting and distributing statistical signature information from shared training data and building generalised models that have satisfactory performance on data from multiple sources. Moreover, concerning the characteristics and contextual information of the ML systems, we offer multiple strategies to perform information extraction adaptively.

Keywords

Edge Computing, Model resilience, Predictive Service, Model Generalization, Distributed Machine Learning, Node Failures

Team:  Qiyuan Wang, Jordi Mateo Fornes, Christos Anagnostopoulos, Kostas Kolomvatsos

Current Outcome

 

Activity

  • Paper presentation: Maintenance of Model Resilience in Distributed Edge Learning Environments [IEEE IE 2023 PaperPresentation], 30 JUNE 2023, 19th IEEE International Conference on Intelligent Environments (IE’23), 27-30 June 2023, Mauritius
  • Work presented in the poster session of the SICSA PhD Conference 2022, Glasgow Caledonian University, 28 and 29 June 2022.
  • Paper to be presented at the IEEE 8th World Forum on Internet of Things, Yokohama, Japan, 26 October–11 November 2022

 Graphical Abstract

Sequential Machine Learning Systems

Distributed ML; Federated/Local Learning; combinatorial Multi-Armed Bandits.

Team: Sham Puthiya, Christos Anagnostopoulos, Rod Murray-Smith, Iadh Ounis

Distributed Detection of Advanced Persistent Threats

Abstract

Since the exposure of SolarWinds attack, the supply chain attacks have emerged in a lot of advanced persistent threats (APTs), which has caused more stealthy attacks. The attackers generally choose target top cited software or repositories, which normally have millions of downstream users. Therefore, the potential influence of a successful supply chain attack is immeasurable. To detect the APTs based on supply chain vulnerabilities in real-time, especially in early stages, it is essential to optimize the detection based on critical stages and also indicators. In addition, the distributed design follows the attribute in spreading of supply chain based attacks. The precise causal correlation between the data and also decisions made in every node contributes to the location of the attack path. It further benefits the reconstruction of the whole attack chain. Through the understanding of the attack path, defenders can quicker take actions to mitigate the ongoing attacks and secure critical digital assets and information.

Keywords Distributed System, Anomaly Detection, Supply Chain Attack, APT, Causality

Team:  Zhuoran Tan, Angelos K.Marnerides, Christos Anagnostopoulos

Graphical Abstract