Machine Learning for Communication Networks

NetSim ML Techniques

AI/ML in 5G RAN

  • NetSim is a leading system level simulator for generating data and training AI/ML algorithms for a wide range of applications that improve RAN performance such as:
    • Traffic steering, Load balancing, Throughput scaling
    • MAC Scheduling, Link adaptation
    • Power control, Beamforming, Interference mitigation
  • NetSim enables rapid creation of a 5G network digital twin and provides support for interfacing AI/ML algorithms for decision making and control in the RAN
    • Verify RAN intelligent Controller (RIC) algorithms
    • AI/ML - NetSim interface for Non-Real Time, Near-Real Time and Real-time RIC feedback loops
    • Synchronize with RAN events and provide AI based decision control at second, milli-second or TTI time scales
    • Equivalent to the O-RAN E2 interface; NetSim simulates the RAN based on your application's (r-app/x-app) control inputs
  • Simulation Application: Online learning for Network Slicing (NetSim in-built, click here for more details)
    • Dynamic resource (PRB) sharing
    • Algorithm: Constrained optimization using Lagrange multipliers
    • Create different slicing scenarios and check if the network meets UE level and slice level SLAs
    • Write your own algorithm by modifying the source code provided with NetSim
  • Simulation Application: Reinforcement learning algorithms - Q-learning, DQN, A2C, PPO and more
    • Connect NetSim to your RL algorithm written in Python
    • Utilize Gymnasium for RL algorithms such as DQN, A2C, PPO, etc.
    • Simulate. NetSim passes the state (e.g: UE measurements, Network measurements, UE context), the RL algorithm passes back the action (e.g: Handover, Power control etc.), and NetSim simulates per the action and returns the rewards (KPIs such as Throughput, Latency etc.) and next state
    • Check how your RL algorithm improves network performance
  • Example 1: Multi-gNB downlink power control to maximize sum throughput (Click here for detailed PDF)
  • Example 2: Load Balancing: 3GPP Release 17 AI/ML for NG RAN Use Case. Framework for Algorithm Development (Click here for detailed article)

AI/ML for Attack Detection in IoT

  • Example: ML Classifier to detect rank attacks in RPL based IoT network (Click here for detailed PDF)
    • Key technical highlights:
      • Implemented RPL based on RFC 6550, including DODAG formation and rank calculations
      • Simulated various attack scenarios with 2 to 15 malicious nodes across network sizes of 6 to 42 nodes
      • Extracted features from packet traces: DAO sent/received, DIO sent/received, and data packets received
      • Normalized features and applied ML classifiers: KNN, Naive Bayes, SVM, and Logistic Regression
    • Results:
      • Confusion matrix generated for each type of classifier
      • Achieved >95% in accuracy, precision and recall
    • Tools used:
      • NetSim for network simulation and data generation
      • Python for data processing, feature extraction, and ML implementation

AI/ML for Attack Detection in VANETs

  • Example: ML Classifier to Detect Sybil Attacks in VANETs (Click here for detailed PDF)
    • Key technical highlights:
      • Implemented VANET scenarios based on IEEE 802.11p and IEEE 1609 standards
      • Simulated attack scenarios with 1 to 3 Sybil nodes across network sizes of 5 to 14 nodes
      • Extracted RSSI-based features: Power, Difference, and Similarity
      • Applied ML classifiers: Random Forest, KNN, XGBoost, and Decision Tree
    • Results:
      • Confusion matrices generated for each classifier type
      • Achieved 95-97% accuracy across classifiers
    • Tools used:
      • NetSim for VANET simulation and data generation
      • Python for data processing, feature extraction, and ML implementation

Generate synthetic data for ML

Machine learning algorithms need ever increasing amounts of data for training and testing. The problems with real data are that it:

  • Can be difficult and expensive to collect. This is particularly true for data that is complex or specialized
  • Is time consuming to label. This process also requires expert knowledge which comes at a high cost.
  • May contain sensitive or confidential information
  • Is often unbalanced; it may not contain enough examples of certain classes or phenomena

On the other hand, synthetic data – that is produced from a simulator - can be:

  • Generated at very low cost and in vast quantities
  • Perfectly labelled, and hence directly used to train neural nets
  • Generated to represent a wide variety of scenarios and edge cases, which can help to improve the robustness and generalizability of machine learning models./li>
  • Created to be free of sensitive or confidential information

Shipped along with are automated utilities for scenario generation and execution; NetSim can generate large amounts of labelled data written to CSV files. The output results and data files generated by NetSim include:

  • Performance Metrics
    • Instantaneous and average throughputs for each link and each application
    • Buffer occupancy vs. time at source and intermediate devices
    • TCP congestion window vs. Time at End each UE and remote Server
  • Packet trace
    • 30+ parameters for every packet as it flows through the network. These include arrival times, queuing times, departure times, payload, overhead, errors, collisions etc.
  • Radio measurements
    • SINR, Pathloss, Shadowing, Fast fading, LOS/NLOS states, O2I Loss, MCS, CQI, BS-UE distances, UE-gNB association.
  • Radio resource allocation
    • Buffer fill (queue size), Scheduling weights, PRBs allocated

Users regularly generate data files that reach up to 10 million rows per simulation. Special cases can cross 100 million rows.

Select list of research publications featuring AI/ML with NetSim

  1. DETONAR: Detection of Routing Attacks in RPL-Based IoT (https://ieeexplore.ieee.org/document/9415869)
  2. Reinforcement-Learning-based IDS for 6LoWPAN (https://ieeexplore.ieee.org/document/9724461)
  3. ELNIDS: Ensemble Learning based Network Intrusion Detection System for RPL based Internet of Things (https://ieeexplore.ieee.org/document/8777504)
  4. Q-Learning Relay Placement for Alert Message Dissemination in Vehicular Networks (https://www.sciencedirect.com/science/article/pii/S1877050922006342)
  5. Adaptive Hybrid Heterogeneous IDS for 6LoWPAN (https://arxiv.org/abs/2205.09170)
  6. Exploring cybersecurity issues in 5G enabled electric vehicle charging station with deep learning (https://ietresearch.onlinelibrary.wiley.com/doi/full/10.1049/gtd2.12275)
  7. Adversarial RL-Based IDS for Evolving Data Environment in 6LoWPAN (https://ieeexplore.ieee.org/document/9916285)
  8. Advancing 6G Network Performance: AI/ML Framework for Proactive Management and Dynamic Optimal Routing (https://ieeexplore.ieee.org/document/10522874)
  9. An Intelligence-Based Framework for Managing WLANs: The Potential of Non-Contiguous Channel Bonding (https://ieeexplore.ieee.org/abstract/document/10500831)
  10. Learning-Based Road Link Quality Estimation for Intelligent Alert-Message Dissemination (https://ieeexplore.ieee.org/abstract/document/10271590)
  11. Malicious Node Detection in VANETs via Enhanced DSR and ML (https://ieeexplore.ieee.org/abstract/document/10532957)
  12. Performance Analysis of 5G DDoS Attack using Machine Learning (https://digitalcommons.memphis.edu/etd/2201/)
  13. SIGMAML: SNR-Guided 5G Mobility Management using Machine Learning Algorithms (https://ieeexplore.ieee.org/abstract/document/10667970)
  14. Intelligent QoS Agent Design for QoS Monitoring and Provisioning in 6G Network (https://ieeexplore.ieee.org/abstract/document/10279078)
  15. A Novel Two-Step Bayesian Hyperparameter Optimization Strategy for DoS Attack Detection in IoT (https://ieeexplore.ieee.org/abstract/document/10467454)
  16. Flexibly Controlled 5G Network Slicing (https://ieeexplore.ieee.org/abstract/document/10019226)

Useful Links