FRIEND: Federated Learning for Joint Optimization of multi-RIS Configuration and Eavesdropper Intelligent Detection in B5G Networks

This paper proposes a privacy-preserving Federated Learning framework that jointly optimizes multi-RIS configuration and eavesdropper detection in B5G cell-free mmWave networks, achieving a 30% improvement in secrecy rate while maintaining high detection accuracy through collaborative DCNN training on local Channel State Information.

The Gist

Imagine a massive, high-tech factory floor where hundreds of robots (the User Equipments) need to talk to each other and to a central control room (the Access Points) instantly and securely. This is the world of B5G (Beyond 5G) and the Industrial Internet of Things (IIoT).

In this paper, the authors are trying to solve a very specific problem: How do we stop spies (eavesdroppers) from listening in on these conversations without slowing everything down or revealing the robots' secrets?

Here is the breakdown of their solution, "FRIEND," using simple analogies.

1. The Setting: A "Cell-Free" Factory

Traditionally, a factory is divided into cells (like rooms), and each robot talks to the nearest wall-mounted speaker. But in this new Cell-Free system, there are no walls. Instead, there are dozens of speakers scattered everywhere, all working together as one giant team.

• The Problem: Because there are no walls, a spy can easily stand anywhere and listen in.
• The Solution (RIS): The authors add Reconfigurable Intelligent Surfaces (RIS). Think of these as smart mirrors placed on the walls. These mirrors can bend and shape radio waves. They can act like a spotlight, beaming the signal only to the intended robot, while making the signal fade into nothingness for anyone standing in the wrong spot (the spy).

2. The Privacy Problem: "Don't Show Your Homework"

To catch the spies, the system needs to learn what a "normal" conversation looks like versus what a "spy" looks like. Usually, you would send all the data from every robot to a central super-computer to train an AI.

• The Risk: Sending all that raw data is like sending your entire homework notebook to a teacher. It's slow, uses up a lot of bandwidth, and risks leaking private secrets if the teacher (or the network) gets hacked.
• The Solution (Federated Learning): The authors use Federated Learning (FL). Imagine instead of sending homework to the teacher, every student (Access Point) solves the problem in their own head. They only send the answer key (the updated AI model) to the teacher, not the homework itself. The teacher combines all the answer keys to create a "Super Smart" global model.
  • Result: The AI gets smarter, but no one ever sees the raw data. Privacy is preserved.

3. The Brain: A "Fast-Track" AI

The AI used to spot the spies is a Deep Convolutional Neural Network (DCNN). It looks at the "Channel State Information" (CSI)—which is basically a complex map of how the radio waves are bouncing around—to decide: "Is this a robot talking, or a spy listening?"

• The Innovation (Early-Exit): Training these AI models is heavy on computer power. To make it faster, the authors added an "Early-Exit" mechanism.
  • The Analogy: Imagine a security guard checking IDs. If the person looks obviously like a robot (99% confidence), the guard lets them through immediately without checking the whole database. Only if the person looks suspicious does the guard do the full, slow check.
  • Result: The system saves a huge amount of time and energy (up to 45% faster) while still catching the spies effectively.

4. The Results: Winning the Game

The authors ran thousands of simulations to see if this "FRIEND" system works.

• Accuracy: The AI became incredibly good at spotting spies, with accuracy reaching up to 95%. It rarely misses a spy (high recall) and rarely mistakes a robot for a spy (low false alarms).
• Security Boost: By using the smart mirrors (RIS) to shape the waves, the system increased the Secrecy Rate (how much secret data can be sent safely) by about 30% compared to systems without these mirrors.
• The Sweet Spot: They found that if you tune the smart mirrors just right (using specific "phase shifts"), you can create a "constructive interference" for the good guys (making their signal strong) and "destructive interference" for the bad guys (making their signal disappear).

Summary

In short, this paper proposes a smart, privacy-friendly security guard for the factories of the future.

1. It uses Federated Learning so the AI learns from everyone without stealing anyone's private data.
2. It uses Smart Mirrors (RIS) to physically block spies from hearing the conversation.
3. It uses a Fast-Track AI that makes quick decisions to save energy.

The result is a network that is faster, more secure, and respects privacy, making it perfect for the next generation of industrial automation.

Technical Summary

Here is a detailed technical summary of the paper "FRIEND: Federated Learning for Joint Optimization of multi-RIS Configuration and Eavesdropper Intelligent Detection in B5G Networks."

1. Problem Statement

The paper addresses the critical security challenges in Beyond 5G (B5G) and 6G Industrial Internet of Things (IIoT) networks, specifically within Cell-Free (CF) millimeter-wave (mmWave) architectures.

• The Challenge: As networks become decentralized and heterogeneous, traditional centralized security mechanisms struggle with scalability, latency, and privacy. Furthermore, the dynamic nature of CF-mMIMO and the susceptibility of mmWave to blockage and fading create complex environments where eavesdropping is a significant threat.
• The Gap: Existing Physical Layer Security (PLS) methods often rely on perfect Channel State Information (CSI) or centralized training, which is impractical due to privacy concerns and the distributed nature of IIoT. There is a lack of unified frameworks that jointly optimize Reconfigurable Intelligent Surface (RIS) configurations and Eavesdropper Detection using privacy-preserving learning.

2. Methodology

The authors propose a novel framework named FRIEND, which integrates Federated Learning (FL), Deep Learning (DL), and RIS technology.

A. System Architecture

• Topology: A Cell-Free mmWave network with distributed Access Points (APs), multiple RIS units, legitimate User Equipments (UEs), and adversarial eavesdroppers.
• Channel Model: The system utilizes 3GPP-compliant channel models (TR 38.901/38.843) for 28 GHz mmWave. The effective channel includes both direct links and RIS-reflected paths.
• Data Generation: A simulation environment (MATLAB) generates realistic CSI data, including multipath fading, blockage, and RIS phase shifts. Data is labeled as "Legitimate" (0) or "Eavesdropper" (1).

B. Federated Learning (FL) Scheme

• Privacy Preservation: Instead of sending raw CSI to a central server, local APs (acting as FL clients) train a Deep Convolutional Neural Network (DCNN) on their local data.
• Aggregation: A central server (located at a central AP) aggregates model weights using the Federated Averaging (FedAvg) algorithm to update the global model. No raw data is exchanged.
• Input Features: The DCNN ingests CSI images (complex-valued channel matrices) fused with metadata (UE transmit power, geolocation coordinates of UE, AP, and RIS).

C. Deep Learning Architecture & Optimization

• Model: A streamlined Deep Convolutional Neural Network (DCNN) designed to process high-dimensional CSI data.
• Early-Exit Mechanism: To address computational constraints and latency, an early-exit branch is added after the second convolutional block. If the model's confidence level (CL) exceeds a threshold (55% or 70%), inference terminates early. This reduces computational load by up to 45% while maintaining accuracy.
• Joint Optimization: The framework evaluates how different RIS phase configurations impact both the Secrecy Rate (SR) and the detection accuracy of the FL model.

3. Key Contributions

1. Unified Framework: Proposes the first integrated approach combining FL, multi-RIS coordination, and deep learning for eavesdropper detection in CF-mmWave IIoT networks.
2. Privacy-Preserving Detection: Demonstrates a method to train robust detection models without sharing sensitive raw channel data, adhering to data privacy regulations.
3. Efficient Inference: Introduces an early-exit mechanism within the FL-DCNN architecture, significantly reducing latency and computational complexity for edge devices without sacrificing detection performance.
4. Systems Engineering Approach: The design follows a rigorous Systems Engineering process (MBSE), defining specific requirements (Availability, Reliability, Scalability, Securability) and validating them through a "System-as-a-Code" simulator.
5. RIS Configuration Impact: Analyzes the trade-off between RIS phase shifts and security performance, showing that specific RIS configurations can actively degrade eavesdropper channels while enhancing legitimate links.

4. Performance Results

Extensive simulations were conducted with 500 UEs (70% legitimate, 30% eavesdroppers), 18 APs, and 3 RIS units.

• Detection Accuracy:
  • The FL-based DCNN achieved Accuracy ranging from 0.71 to 0.93 (interquartile center ~0.84).
  • Recall (critical for PLS to avoid missing attackers) reached up to 0.95, with high F1-scores indicating a balance between precision and recall.
  • The model successfully generalized across different channel conditions despite data heterogeneity.
• Computational Efficiency (Early-Exit):
  • Implementing the early-exit mechanism reduced inference time by 35% (at 70% confidence) and 45% (at 55% confidence).
  • Accuracy remained competitive (>0.83) even with early termination.
• Secrecy Rate (SR) Improvement:
  • The integration of FL and optimal multi-RIS coordination improved the Average Secrecy Rate (ASR) by approximately 30% compared to baseline non-RIS-assisted methods.
  • Specific RIS phase configurations (e.g., Phase ID 89) achieved SRs exceeding 20 bps/Hz in high legitimate-to-eavesdropper ratios, significantly outperforming non-ML approaches.

5. Significance and Future Work

• Significance: This work establishes a scalable, distributed, and privacy-preserving paradigm for Physical Layer Security in next-generation industrial networks. It proves that intelligent, data-driven approaches can effectively counter eavesdropping in complex, decentralized mmWave environments without compromising user privacy.
• Future Directions:
  • Validation using real-world data to fulfill operational requirements.
  • Evaluation of the model under scalable and alternative RIS configurations.
  • Investigation of the early-exit mechanism in offloading tasks to further optimize edge computing resources.

In conclusion, the FRIEND framework successfully bridges the gap between advanced wireless physical layer technologies (RIS, mmWave) and modern AI paradigms (FL, Deep Learning), offering a robust solution for secure, high-capacity B5G/6G IIoT deployments.