🔢 mathematics

Deep Reinforcement Learning for Cognitive Time-Division Joint SAR and Secure Communications

This paper proposes a deep reinforcement learning-based dynamic time-division framework that integrates cognitive synthetic aperture radar with secure communications to jointly optimize time and power allocation, thereby maximizing secrecy rates against moving eavesdroppers while satisfying sensing constraints.

Original authors: Mohamed-Amine Lahmeri, Ata Khalili, Yujiao Liu, Anke Schmeink, Robert Schober

Published 2026-04-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohamed-Amine Lahmeri, Ata Khalili, Yujiao Liu, Anke Schmeink, Robert Schober

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a pilot flying a high-tech drone over a city. Your mission has two jobs:

Talk to a friendly ground station (a user) to send them secret messages.
Watch out for a sneaky spy on the ground who is trying to steal those messages.

The problem? The spy is moving. They might be in a car, running, or hiding behind buildings. If you don't know exactly where they are, you can't protect your secret messages. Traditional methods are like guessing where the spy is, which often fails.

This paper proposes a brilliant new strategy called "Cognitive Time-Division Joint SAR and Secure Communications." That's a mouthful, so let's break it down with some simple analogies.

The Core Idea: The "Look-Then-Act" Dance

Think of your drone as a magical lighthouse that can switch between two modes: Scanning and Shouting.

The Scanning Mode (SAR):
Instead of just looking with eyes, the drone uses a super-powerful radar (called Synthetic Aperture Radar, or SAR) to take a "mental snapshot" of the ground. It's like using a high-speed camera to take a picture of a moving car.
- The Trick: The drone uses a technique called Along-Track Interferometry (ATI). Imagine the drone has two "ears" (antennas) spaced apart. As it flies, it listens to the spy's reflection with both ears. Because the spy is moving, the sound (or signal) hits the ears at slightly different times. This allows the drone to calculate exactly where the spy is and how fast they are moving, even if they are trying to hide in the "static" noise of the city.
The Shouting Mode (Communication):
Once the drone knows where the spy is, it switches to "Shouting." It aims a laser-sharp beam of data directly at the friendly user.
- The Defense: To make sure the spy can't hear the message, the drone also blasts "white noise" (artificial jamming) directly at the spy's last known location. It's like shouting a secret to your friend while simultaneously playing a loud, confusing siren right next to the eavesdropper's ear.

The Big Challenge: The Balancing Act

Here is the tricky part: The drone has a limited amount of battery and time.

If it spends too much time scanning, it misses the chance to send messages.
If it spends too much time shouting, it loses track of the moving spy, and the spy might steal the secrets.

It's like a chef trying to bake a cake while also watching a pot of boiling water. If they stare at the pot too long, the cake burns. If they focus only on the cake, the water boils over.

The Solution: The "Smart Brain" (Deep Reinforcement Learning)

The authors realized that a human pilot couldn't calculate the perfect split-second timing for every single situation, especially if the spy suddenly speeds up or changes direction.

So, they gave the drone a Super-Brain powered by Deep Reinforcement Learning (DRL).

How it learns: Imagine the drone is a video game character. It plays the game thousands of times against different types of spies (some slow, some fast, some zig-zagging).
The Reward: Every time it successfully sends a secret message without the spy stealing it, it gets a "point." Every time the spy steals a message or the drone wastes time, it loses points.
The Result: The drone learns a perfect strategy. It learns to scan more when the spy is running fast (because they move out of the "safe zone" quickly) and to talk more when the spy is slow or far away.

Why This Paper is a Big Deal

It's "Cognitive": The system isn't just following a rigid rulebook. It adapts in real-time. If the spy speeds up, the drone instantly knows, "Okay, I need to scan more often," and adjusts its schedule automatically.
It's Robust: Even if the spy does something the drone has never seen before (like a weird zig-zag pattern), the "Super-Brain" is smart enough to figure out the best move on the fly.
It Solves the "Clutter" Problem: In a city, there are lots of stationary things (buildings, trees) that confuse normal radar. This system is specifically designed to ignore the "static" background and only focus on the "moving" spy.

The Bottom Line

This paper presents a system where a drone acts like a smart security guard. It constantly checks its surroundings to find a moving thief, and the moment it spots the thief, it instantly changes its behavior to protect the secret conversation. It uses a "learning" algorithm to figure out the perfect balance between watching and talking, ensuring that even the fastest, sneakiest spy can't steal the secrets.

In short: It's a drone that learns to dance with a moving target to keep secrets safe.

1. Problem Statement

The paper addresses the critical challenge of securing Air-to-Ground (A2G) communications against ground-moving eavesdroppers.

The Challenge: Conventional physical-layer security methods (like beamforming and jamming) rely on accurate Channel State Information (CSI) or prior knowledge of the eavesdropper's location. In dynamic scenarios (e.g., surveillance or post-disaster), the eavesdropper's position and velocity are often unknown, and ground clutter makes detection difficult for standard radar.
The Gap: While Synthetic Aperture Radar (SAR) can detect moving targets, integrating it with communication systems (Joint SAR and Communication - JSARC) is complex. Existing works often assume static environments or oversimplify the coexistence of sensing and communication, failing to handle dynamic, moving threats effectively.
The Goal: To design a framework that dynamically allocates time and power between Sensing (SAR) and Communication to maximize the worst-case secrecy rate while tracking a moving eavesdropper and satisfying quality-of-service constraints.

2. Methodology

The authors propose a Cognitive Time-Division Joint SAR and Communication (TD-JSARC) framework enhanced by Deep Reinforcement Learning (DRL).

A. System Model

Scenario: An Aerial Base Station (ABS) equipped with a Uniform Linear Array (ULA) serves a ground User Equipment (UE) while a ground-moving eavesdropper attempts to intercept the signal.
Sensing Mechanism: The ABS uses Along-Track Interferometry (ATI) SAR. Two specific antennas are dedicated to sensing to separate moving targets from ground clutter.
- Trade-off: Longer sensing intervals improve localization accuracy (resolution) but reduce time available for data transmission.
Uncertainty Modeling: Since the eavesdropper's trajectory is unknown, the system maintains an uncertainty region (a circular disk) around the last estimated position. This region expands over time based on the eavesdropper's maximum speed and acceleration limits.
Communication Strategy:
- Beamforming: Maximum Ratio Transmission (MRT) is used to maximize the signal strength at the legitimate UE.
- Jamming: Artificial Noise (AN) is generated and directed toward the last estimated eavesdropper position to degrade the eavesdropper's signal-to-noise ratio (SNR).

B. Optimization Problem

The core problem is a non-convex mixed-integer optimization problem aimed at maximizing the average worst-case secrecy rate.

Variables:
- Discrete: Number of time frames (apertures), duration of each frame, and duration of sensing vs. communication within each frame.
- Continuous: Power-splitting factor ( $\alpha$ ) between information signal and artificial noise.
Constraints: Minimum Signal-to-Clutter Ratio (SCR) for SAR detection and minimum average data rate for the UE.

C. Solution via Deep Reinforcement Learning (DRL)

Due to the unknown eavesdropper trajectory and the complexity of the mixed-integer problem, the authors formulate the task as a Markov Decision Process (MDP) and solve it using Proximal Policy Optimization (PPO).

State Space ( $S$ ): Includes current aperture length, last estimated eavesdropper velocity, current uncertainty radius, angular separation between the eavesdropper and user, and relative distance differences.
Action Space ( $A$ ): A binary decision at each time slot: Sensing ( $a[n]=1$ ) or Communication ( $a[n]=0$ ).
Reward Function ( $R$ ): Designed to maximize secrecy rate while penalizing violations of SCR and minimum rate constraints.
Hybrid Approach:
- The DRL agent handles the discrete time-scheduling (when to sense vs. communicate).
- The continuous power allocation ( $\alpha$ ) is solved separately via a low-dimensional robust optimization (exhaustive grid search) based on the current state, ensuring the solution remains computationally efficient.

3. Key Contributions

Novel Framework: Proposes the first cognitive TD-JSARC framework specifically designed to mitigate ground-moving eavesdroppers using SAR ATI.
Dynamic Uncertainty Model: Develops a model that dynamically updates the eavesdropper's uncertainty region based on SAR resolution and kinematic constraints (max speed/acceleration), allowing the system to adapt to unknown trajectories.
MDP Formulation: Successfully reformulates a complex, non-convex mixed-integer resource allocation problem into an MDP, enabling the use of DRL for real-time decision-making without prior trajectory knowledge.
Generalization: The proposed learning-based solution demonstrates the ability to generalize to previously unseen eavesdropper motion patterns, a significant advantage over static or rule-based schemes.

4. Simulation Results

The authors validated the framework using extensive simulations with parameters typical for UAV-based systems (e.g., 1 GHz radar bandwidth, 100m altitude).

Performance vs. Benchmarks: The proposed PPO-based method significantly outperforms three baseline schemes:
1. Hybrid DRL: A scheme attempting to learn both discrete and continuous variables jointly (which proved less stable/efficient).
2. Equal-Aperture: A non-learning scheme with fixed sensing/communication ratios.
3. Random Allocation: Random time allocation satisfying constraints.
Impact of Speed: As the eavesdropper's speed increases, the performance gap widens. The proposed method adapts by increasing the Sensing-to-Communication (S2C) ratio during high-speed intervals to maintain tracking accuracy, whereas static schemes fail to adapt.
Secrecy Rate: At an eavesdropper speed of 14 m/s, the proposed scheme achieved an average worst-case secrecy rate of ~3.3 bits/s/Hz, compared to ~2.7 bits/s/Hz for the best benchmark.
Robustness: The system successfully maintained constraints (SCR and minimum rate) even when tested on linear trajectories with oscillating speeds that were not present during the training phase.

5. Significance

This work bridges the gap between cognitive radar and secure communications. It demonstrates that:

Sensing can enable security: By actively sensing the environment, a system can dynamically adjust its communication strategy to counter moving threats, rather than relying on static assumptions.
AI is essential for dynamic JSARC: Traditional optimization methods struggle with the unknown, time-varying nature of moving eavesdroppers. DRL provides a robust, adaptive solution that learns optimal scheduling policies in real-time.
Practical Applicability: The framework is relevant for critical scenarios like military surveillance, disaster relief, and secure UAV operations where adversaries are mobile and their locations are initially unknown.

The paper concludes that the proposed cognitive, learning-based approach offers a superior balance between sensing accuracy and communication security compared to existing static or non-learning methods.