AI-powered eVolution towards opEn and secuRe edGe architEctures

Project Information

VERGE

Grant agreement ID: 101096034

DOI

10.3030/101096034

Project closed

EC signature date 22 November 2022

Start date 1 January 2023

End date 30 June 2025

Funded under

• Digital, Industry and Space

Total cost

€ 5 263 796,86

EU contribution

€ 4 898 437,50

Coordinated by

UNIVERSITAT POLITECNICA DE CATALUNYA
Spain

Periodic Reporting for period 2 - VERGE (AI-powered eVolution towards opEn and secuRe edGe architEctures)

Reporting period: 2024-01-01 to 2025-06-30

Summary of the context and overall objectives of the project

The main goal of VERGE has been to provide an integrated approach on how to tackle the challenges of edge computing evolution described around three main pillars: (1)“Edge for AI (Edge4AI)”, namely a flexible, modular and converged edge platform design, unifying the Life Cycle Management (LCM) and closed-loop automation for cloud-native applications, Multi-access Edge Computing (MEC) and network services across a multi-domain edge-cloud continuum, while fully exploiting the capabilities of cutting-edge multi-core and multi-accelerator platforms for ultra-high computational performance. (2) “AI for edge (AI4Edge)” namely an AI-powered portfolio of solutions that will leverage the multitude of information and metrics provided by the VERGE platform monitoring mechanisms to manage and orchestrate the VERGE platform computing and network resources. (3) “Security, privacy and trustworthiness for AI (SPT4AI)”, through a suite of methods and tools that will ensure a) security of the AI-based models against adversarial attacks, b) privacy of data and models, and c) sample-efficient training, safety verification and explainability for model decisions, resulting in improved trust in AI models.
This main goal has been achieved through the following specific obejctives:
1.- Design an open and secure edge computing architecture to efficiently support highly demanding XR and IoT-driven applications in terms of computation and connectivity, relying on an edge-cloud continuum that fosters virtualisation of massive heterogeneous edge computing resources across a multi-access RAN.
2.-Design and build the Edge4AI concept that provides the infrastructure, mechanisms and programming tools to enable the development, deployment and orchestration of massively distributed AI processes across heterogeneous computing and memory resources.
3.-Design an AI4Edge framework that encompasses cutting-edge AI solutions for managing and orchestrating the edge computing and the RAN resources towards an optimum performance that satisfies the highly demanding applications requiring extremely low latencies and/or very high-capacity justifying edge processing and computing.
4.-Develop tools that ensure the security, privacy and trustworthiness of the VERGE system.
5. Showcase the VERGE solutions by means of Proof of Concept (PoC) demonstrations.
6. Carry out extensive dissemination, standardisation and exploitation activities.

The project solutions have impacted on different improvement areas, namely reduced latency, enhanced service availability, enhanced service reliability, efficient resource utilization, energy efficiency, flexibility/scalability, enhanced computation, efficient AI model training, AI safety and AI explainability.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

• The consolidation of the VERGE system architecture, integrating all key architectural blocks and internal components to deliver an integrated, B5G-enabled, multi-site edge-cloud compute continuum (Edge4AI), supporting intelligent (AI4Edge) and secure (SPT4AI) AI/ML pipelines.
• The consolidation of the AI/ML LCM, with intelligent AI4Edge solutions and SPT4AI trustworthiness mechanisms seamlessly integrated into a unified software pipeline.
• The consolidated design of the VERGE open dataspace has been successfully completed. A key achievement is the development of a standardized schema encompassing dataset definitions, metadata specifications, version control mechanisms, and governance policies.
•Accelerated and distributed computing capabilities, coupled with advanced learning paradigms.
• AI-driven hierarchical and modular orchestration, management, and control.
• A portfolio of AI-driven solutions for network slicing and resource optimization.
• AI trustworthiness solutions, addressing security, privacy, robustness, safety, and explainability challenges.
• Unified data access and LCM of all AI-driven pipelines through the VERGE Cognitive Framework.
• The planning, integration, and validation of the final project demonstrations, resulting in the production of seven PoC demonstrations by the end of the project.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

• Optimal split point selection for complex Deep Neural Networks (DNNs), achieving a 19% reduction in inference time (compared to full computation on the User Equipment (UE)), and a 9.5% reduction in energy consumption at the UE.
• Reduction of up to 79% in the lines of code required to create distributed workflows, significantly enhancing programmability.
•Up to 45% power saving has be achieved by efficient resource allocation and function offloading leveraging FPGA programmability.
• Time-triggered Federated Learning (FL) methodologies ensuring up to 40% reduction in communication cost and faster convergence in resource-constrained environments.
•37% end-to-end latency reduction when the MEC app slice concept is employed to prioritize critical services.
• 36–97% latency reduction has been achieved through autoscaling capabilities of the distributed workflow orchestration (depending on the number of available computing resources), as well as through regression-based approaches for intent-driven edge-site resource allocation and rightsizing.
• Optimal MEC selection and network autoscaling, yielding up to 28% latency reduction and 98.5% service reliability without compromising utilization efficiency.
• Relay management and optimized low-layer functional split selection, leading to up to 23% cost reduction (compared to fixed-splitting solutions).
• Enhanced and generalizable attention-based mechanism for forecasting edge-cloud resource utilization, resulting in 59.4% and 93.4% improvements in prediction accuracy and inference speed, respectively, which could be leveraged by the multi-site orchestrator.
• Homology-based solutions leveraging semantic transmission achieving a 15-fold improvement in resource utilization.
• By employing transfer learning techniques, up to 80% training time reduction was achieved in a Deep Reinforcement Learning (DRL)-based network slicing solution when RAN topologies were expanded.
• Security mitigations for two AI-driven tasks (beam selection and power control) in a Distributed Multiple Input Multiple Output (D-MIMO) environment, decreasing the attack success rate to below 1% and mitigating poisoning attacks while maintaining data validity.
• A selective knowledge transfer approach reducing the attack success rate to 7% in collaborative hazard detection scenarios in untrusted transportation environments.
• A homomorphic encryption approach to avoid exposing private data during model aggregation reduced the leakage attack success rate to less than 10%, while maintaining model accuracy within a 3% margin.
• A framework for trustworthy RL was proposed, ensuring reliable decision-making with a strong trade-off between accuracy (92%) and robustness (94%) under previously unseen conditions.
• Causal RL methodologies demonstrated over 40% better sample efficiency than traditional methods and improved generalization.
• Formal verification framework enabling probabilistic estimation of whether key safety, reliability, and resilience properties are fulfilled.
• AI feature explainability for DRL models was supported through a Shapley Value-based approach.
• 75.6% reduction in total pipeline time for training and validation of an ML model was achieved through CF services, thanks to enhanced pipeline orchestration and parallel execution