Objective
This axis aims to bring together work exploring the dual interaction between Artificial Intelligence and communication networks:
- On one hand, using AI to optimize, automate, or secure networks;
- On the other, treating networks as critical infrastructure for training and executing distributed AI models.
The goal is to foster interdisciplinary collaboration around these two complementary themes, with research spanning from models and algorithms to implementation on experimental platforms (Kubernetes, 5G testbeds, edge computing, IoT, etc.).
Theme 1: AI for Networks
This theme focuses on the contribution of AI to make networks more intelligent, adaptive, autonomous, secure, and efficient.
Examples of research directions in this area include (but are not limited to):
- Autonomous control and network optimization:
- Example methods: Markov Decision Process, Proximal Policy Optimization, Deep Q-Learning, Metaheuristics for routing, slicing management, resource allocation…
- Applications: Dynamic placement of network functions (VNF), scheduling in Kubernetes clusters…
- Supervised learning for anomaly detection or traffic classification:
- Example methods: Random Forests, SVMs, CNNs/1D for time sequences.
- Applications: DDoS attack detection, encrypted application recognition…
- Unsupervised or self-supervised learning for network monitoring:
- Example methods: Autoencoders, clustering (k-means, DBSCAN), contrastive learning…
- Applications: Emerging behavior identification, monitoring in 6G environments…
- Graph Neural Networks (GNNs) to capture network topology:
- Example methods: PyTorch Geometric, DGL…
- Applications: Congestion prediction, predictive routing, Virtual Network Embedding (VNE)…
Theme 2: Networks for AI
This theme investigates how networks can support the growing demands of AI applications, which are often computation- and traffic-intensive, distributed, interactive, or embedded.
Key research directions include:
- Federated and decentralized learning on constrained networks:
- Example methods: FedAvg, SCAFFOLD, FedNova, quantization and sparsity techniques…
- Applications: Asynchronous synchronization, adaptive client selection, gradient compression.
- Networks for real-time AI (robotics, edge, autonomous vehicles):
- Example technologies: URLLC, TSN (Time-Sensitive Networking), ultra-low latency protocols…
- Applications: Drone coordination, real-time detection in robotics.
- Optimization of model and data transport for AI:
- Example methods: Model splitting for transformer-based models (e.g., BERT on edge/cloud), dynamic offloading via MDP.
- Techniques: Network-Aware Model Partitioning, Multi-access Edge Computing (MEC).
- Network programmability to support AI:
- Examples: SDN/OpenFlow for dynamic AI flow control; P4 for intelligent dataplane processing.
Theme 3: Co-evolution of AI ↔ Networks
This theme explores feedback loops between networks and AI for optimized co-design:
- In-network learning: Embedded learning inside the network (e.g., using SmartNICs, eBPF suite).
- Co-optimized simulation: Training AI models within network simulators (e.g., ns-3 + reinforcement learning).
- Joint design of AI and network policies: Co-training of models and scheduling/routing policies.
Axis Coordinators:
- Guillaume LOZENGUEZ (IMT Nord-Europe)
- Badii JOUABER (Télécom SudParis)
