Enhancing PLS of Indoor IRS-VLC Systems for Colluding and Non-Colluding Eavesdroppers

This paper proposes a deep reinforcement learning-based approach using proximal policy optimisation to enhance physical layer security in indoor visible light communication systems by leveraging realistic IRS-induced time delays to constructively boost signals for legitimate users while intentionally creating intersymbol interference for both colluding and non-colluding eavesdroppers.

The Gist

Imagine you are in a room trying to have a secret conversation with a friend (let's call him Bob) using a very bright, fast-blinking flashlight (the LED). The light carries your secret message.

The problem? The room is full of other people (Eavesdroppers or "Eves") who are also looking at the light. Since light travels in straight lines and bounces off walls, these spies can easily catch a glimpse of your message, even if they aren't standing right next to Bob.

Traditionally, people tried to solve this by making the flashlight beam very narrow or by using complex math codes (encryption) to scramble the message. But this paper proposes a clever, physical trick using a high-tech mirror wall called an IRS (Intelligent Reflecting Surface).

Here is the simple breakdown of how this works, using everyday analogies:

1. The Problem: The "Echo" Effect

In a normal room, light hits the walls and bounces back. Usually, we think of this as just "noise" or a delay. If you shout in a canyon, you hear your voice bounce back a split second later. In high-speed data, these tiny delays can mess up the signal, causing "Inter-Symbol Interference" (ISI)—basically, the words of your message start to blur together.

Most security systems ignore this blurring. This paper says: "Let's use the blur as a weapon!"

2. The Solution: The "Smart Mirror Wall"

Imagine the wall has hundreds of tiny, adjustable mirrors (the IRS elements). The researchers figured out how to tilt these mirrors so that the light bounces off them and arrives at the receiver at very specific times.

• For Bob (The Friend): The mirrors are tilted so that the bouncing light arrives at the exact same moment as the direct light from the flashlight.

  • Analogy: Imagine a choir. The direct voice is the lead singer. The mirrors are the backup singers. The researchers tune the mirrors so the backup singers hit their notes perfectly in sync with the lead. The result? A massive, powerful, clear sound. Bob's signal gets stronger and clearer.

• For the Spies (The Eavesdroppers): The mirrors are tilted so that the bouncing light arrives at the wrong time compared to the direct light.

  • Analogy: Imagine the backup singers are now shouting out of sync, hitting the wrong notes, and drowning out the lead singer. The sound becomes a chaotic mess of noise. The spies' signal becomes garbled and useless.

3. The Two Types of Spies

The paper tests this against two types of bad guys:

• Non-Colluding Spies: They act alone. They can't talk to each other. The system just needs to make sure none of them can hear clearly.
• Colluding Spies: This is the scary scenario. They are like a team of spies sharing their notes. If Spy A hears a little bit and Spy B hears a little bit, they combine their notes to figure out the whole message.
  • The Paper's Win: Even when the spies are working together and are physically closer to the flashlight than Bob is (a "worst-case" scenario), the smart mirror wall still manages to make Bob's signal clear and the spies' signal a mess.

4. The "Brain" Behind the Mirrors (AI)

You might ask: "How do you know exactly how to tilt hundreds of tiny mirrors to make this happen?"
There are too many combinations to try them all one by one (it would take longer than the age of the universe).

So, the authors used a Robot Brain (AI) called Proximal Policy Optimization (PPO).

• Analogy: Think of a video game character trying to solve a maze. At first, the character runs into walls and fails. But every time it tries a new path, it gets a "score." Over thousands of tries, the character learns the perfect path without needing to be told exactly how to move.
• In this case, the AI "plays" with the mirror angles. It tries different combinations, sees which one makes Bob's signal the strongest and the spies' signal the weakest, and learns the perfect setting instantly.

5. The Results: A Huge Win

The simulation showed that this method is incredibly powerful:

• In the best scenarios, Bob's ability to receive the secret message improved by 67%.
• In the worst scenarios (where the spies are closer and stronger than Bob), the AI's mirror arrangement turned a losing game into a winning one. It improved the security by 107% (for teaming spies) and 235% (for lone spies) compared to just giving all the mirrors to Bob.

Summary

This paper is about turning a physical weakness (light bouncing off walls and causing delays) into a superpower. By using a smart, AI-controlled mirror wall, we can create a "soundproof room" for light. We make the signal a perfect symphony for our friend and a chaotic noise storm for the spies, keeping our secrets safe even if the spies are standing right next to us.

Technical Summary

Here is a detailed technical summary of the paper "Enhancing PLS of Indoor IRS-VLC Systems for Colluding and Non-Colluding Eavesdroppers."

1. Problem Statement

The paper addresses the security challenges in Indoor Visible Light Communication (VLC) systems, specifically focusing on Physical Layer Security (PLS).

• The Gap: Most existing research on Intelligent Reflecting Surface (IRS)-aided VLC assumes frequency-flat channels, ignoring the time delays introduced by reflected paths. This assumption is valid for narrowband systems but fails in practical high-speed, wideband VLC systems (e.g., using DCO-OFDM) where path length differences cause Inter-Symbol Interference (ISI) and frequency-selective fading.
• The Threat: Indoor VLC signals are confined by line-of-sight (LoS) but can be intercepted by any user within the illumination footprint. The paper considers two threat models:
  1. Non-colluding Eavesdroppers: Act independently.
  2. Colluding Eavesdroppers: Cooperate by combining their received signals (e.g., via Maximal Ratio Combining), creating a significantly stronger eavesdropping channel.
• The Challenge: The goal is to maximize secrecy capacity (the difference between the legitimate user's rate and the eavesdropper's rate) by manipulating the IRS. However, the problem involves discrete allocation of IRS elements, making it a non-convex, high-dimensional combinatorial optimization problem that is computationally intractable for exhaustive search.

2. Methodology

The authors propose a framework that leverages IRS-induced time delays as a degree of freedom to shape the channel frequency response.

A. System Model

• Scenario: A single LED transmitter (Alice) serves a legitimate user (Bob) in the presence of $K$ passive eavesdroppers ($E_1, \dots, E_K$). An IRS with $N_{irs}$ passive elements is mounted on a wall.
• Channel Modeling: The channel is modeled as a superposition of the direct LoS path and multiple Non-Line-of-Sight (NLoS) paths reflected by the IRS.
  • The Channel Impulse Response (CIR) accounts for distinct time delays ($\tau$) based on the geometric distances of the LoS and reflected paths.
  • The Channel Frequency Response (CFR) is derived via Fourier transform, showing that the magnitude squared depends on cosine terms involving delay differences. This allows for constructive or destructive interference at specific frequencies.
• Secrecy Capacity Formulation:
  • Non-colluding: Secrecy capacity is limited by the eavesdropper with the highest individual rate.
  • Colluding: Eavesdroppers use Maximal Ratio Combining (MRC), summing their individual SNRs. The secrecy capacity is the difference between Bob's rate and the combined rate of the eavesdroppers.

B. Optimization Problem

The objective is to find an optimal discrete allocation vector $\mathbf{a}$, where each IRS element is assigned to serve either Bob or one of the eavesdroppers.

• Goal: Maximize $C_s(\mathbf{a})$ (Secrecy Capacity).
• Constraint: Each element must be allocated to exactly one user.
• Complexity: The search space grows exponentially as $(K+1)^{N_{irs}}$, rendering exhaustive search impossible for practical IRS sizes (e.g., $15 \times 15$ elements).

C. Proposed Solution: Deep Reinforcement Learning (PPO)

To solve the combinatorial problem, the authors formulate it as a Markov Decision Process (MDP) and solve it using Proximal Policy Optimization (PPO).

• Agent: The IRS controller.
• State ($s_t$): The history of allocations made for previous elements ($a_1, \dots, a_{t-1}$) combined with user coordinates.
• Action ($a_t$): Assigning the current $t$-th IRS element to a specific user ($B, E_1, \dots, E_K$).
• Reward ($R$): The terminal secrecy capacity calculated after all elements are allocated. Intermediate rewards are zero (sparse reward).
• Mechanism: The PPO algorithm learns a policy to allocate elements such that reflected signals align constructively with the LoS path at Bob (boosting his rate) while creating destructive interference (ISI) at the eavesdroppers (degrading their rate).

3. Key Contributions

1. Realistic Time-Delay Modeling: The paper is among the first to explicitly model and exploit IRS-induced time delays in wideband indoor VLC for PLS, moving beyond the simplified frequency-flat assumption.
2. Unified Threat Model: It provides a comparative analysis of secrecy capacity under both colluding and non-colluding eavesdropper scenarios within a unified time-delay IRS framework.
3. PPO-Based Optimization: The authors develop a novel PPO-based RL algorithm to solve the discrete IRS allocation problem. This approach avoids the exponential complexity of exhaustive search and learns an effective allocation strategy through interaction.
4. Analytical Approximation: A closed-form approximation for the achievable rate is derived to facilitate analysis, validated against exact numerical integration.

4. Simulation Results

Simulations were conducted in a $5m \times 5m \times 3m$room with a$15 \times 15$ IRS array.

• Best-Case Geometry (Bob near LED, Eavesdroppers far):

  • The PPO-optimized allocation increased Bob's achievable rate by ~67% compared to LoS-only.
  • Eavesdropper rates were reduced by ~6%.
  • Secrecy capacity improved significantly over a baseline where all IRS elements were allocated to Bob ("All-Bob").

• Worst-Case Geometry (Bob at cell edge, Eavesdroppers near LED):

  • In this scenario, eavesdroppers naturally have stronger channels than Bob.
  • Colluding Case: The PPO approach improved secrecy capacity by 107% compared to the "All-Bob" baseline, successfully turning a negative secrecy capacity (insecure) into a positive one.
  • Non-Colluding Case: The improvement was even more dramatic, reaching 235% over the baseline.
  • Bob's rate increased by ~192% (colluding) and ~191% (non-colluding) compared to LoS-only, while eavesdropper rates were suppressed.

• Convergence: The PPO algorithm converged quickly (within a few hundred episodes) and consistently outperformed static baselines across different IRS sizes ($4 \times 4$and$15 \times 15$).

5. Significance

This work demonstrates that time-delay-based IRS control is a powerful tool for securing indoor VLC systems.

• Robustness: The proposed method ensures secure communication even in geometrically disadvantageous scenarios where eavesdroppers are physically closer to the transmitter than the legitimate user.
• Practicality: By using PPO, the solution offers a computationally feasible way to manage complex, wideband IRS configurations in real-time, addressing a critical gap in current PLS literature which often ignores wideband effects.
• Security Paradigm: It shifts the security strategy from merely blocking signals to engineering the channel impulse response, using the inherent physics of wideband propagation (ISI) as a weapon against eavesdroppers.