⚡ electrical engineering

Diffusion Fluid Antenna Systems for Resilient ISAC

This paper proposes "Diffusion Fluid Antenna Systems," a generative AI-driven framework that leverages a conditional denoising diffusion probabilistic model to optimize fluid antenna port selection for resilient Integrated Sensing and Communication, enabling capabilities such as generative spatial stealth and target isolation by dynamically reshaping sensing signatures on the electromagnetic fading manifold.

Original authors: Noor Waqar, Kai-Kit Wong, Chan-Byoung Chae, Ross Murch

Published 2026-05-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Noor Waqar, Kai-Kit Wong, Chan-Byoung Chae, Ross Murch

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

The Big Picture: A New Way to "See" and "Hide"

Imagine a world where your phone or a smart device doesn't just talk to a tower (communication) but also acts like a radar to "see" its surroundings (sensing). This is called ISAC (Integrated Sensing and Communication).

Usually, engineers try to make this work better by tweaking the signal itself—like changing the volume or the rhythm of the radio waves. But this paper argues that's not enough. Instead, they propose changing the shape of the antenna itself.

Think of a standard antenna like a rigid, fixed-size mirror. You can't change its shape. This paper introduces a Fluid Antenna System (FAS). Imagine a mirror made of liquid mercury or a flexible sheet of thousands of tiny, movable pixels. You can instantly reshape it, moving the "active" parts to different spots to catch or block signals in specific ways.

The authors use a special type of Generative AI (called a "Diffusion Model") to figure out exactly how to reshape this liquid antenna to either hide a device from radar or clear away interference so a radar can see a target clearly.

The Two Main Tricks (Use Cases)

The paper demonstrates two specific "superpowers" this system can achieve:

1. The "Ghost" Mode (Stealth)

The Problem: Sometimes, you don't want to be seen. Maybe you are a drone that needs to stay hidden, or a device that shouldn't trigger a radar alarm.
The Old Way: You might try to jam the radar (send noise), but that's like shouting to hide a whisper—it pollutes the whole room and is often illegal or obvious.
The New Way (Diffusion-FAS): The device uses its fluid antenna to "morph" its shape. It moves its active parts to spots where the radio waves naturally cancel each other out (like finding a "dead zone" in a room where sound doesn't travel).
The Result: The device becomes statistically invisible. It looks exactly like background static or "clutter" (like rain or dust) to the radar. The paper claims this can make the device 100 times harder to detect (a drop of two orders of magnitude) without sending any jamming signals.

2. The "Spotlight" Mode (Target Isolation)

The Problem: Imagine a radar trying to find a specific car (Target A). But right next to it is a noisy truck (Target B) that is reflecting so much signal that it blinds the radar, making it impossible to see the car.
The Old Way: The radar struggles to separate the two because they are too close together.
The New Way (Diffusion-FAS): The noisy truck (User B) is also equipped with a fluid antenna. The truck's AI calculates a shape that creates a "shadow" or a "hole" in its own reflection specifically aimed at the radar's view of the car.
The Result: The truck effectively "turns off" its reflection in the direction of the car. It creates a spatial null (a silence zone) that lets the radar see the car clearly, even though the truck is right next to it.

How the AI Works: The "Denoising" Analogy

The hardest part of this is that the device doesn't know the full map of the radio waves; it only sees a few scattered clues (sparse measurements). How does it know where to move the antenna?

The authors use a Denoising Diffusion Probabilistic Model (DDPM). Here is a simple analogy:

Imagine you have a picture of a perfect, clear sky (the ideal antenna shape), but someone has covered it with thick, swirling fog (noise).

The Forward Process: The AI starts with a clear image and slowly adds more and more fog until it's just white noise.
The Reverse Process (What the AI does): The AI is trained to look at a foggy picture and guess what the clear image underneath should look like. It peels back the fog layer by layer.

In this paper, the "fog" is the uncertainty of the radio environment. The AI starts with a random guess of where to put the antenna ports and then "peels back the fog" using its training. It uses a special "energy guide" (a mathematical rule) to nudge the guess toward the perfect shape that either hides the user or clears the interference.

Why is this special?
Usually, finding the perfect antenna shape is a math nightmare (an "NP-hard" problem) because there are billions of combinations. Trying them all is impossible. This AI doesn't try them all; it "dreams" the best solution by learning the patterns of how radio waves behave, allowing it to find the perfect shape in real-time.

What the Experiments Showed

The authors ran simulations to test their idea. Here is what they found, sticking strictly to their results:

Small is Mighty: You don't need a massive antenna. Even a very small, compact fluid antenna (about the size of a few wavelengths) was enough to create these "ghost" or "spotlight" effects.
Few Clues are Enough: The AI could figure out the perfect antenna shape even if it only "saw" a tiny fraction (less than 10%) of the available antenna ports. It filled in the rest of the picture using its learned knowledge.
Beating the Competition:
- Random Selection: If you just randomly pick antenna spots, it doesn't work well.
- Static Antennas: If you use a normal, fixed antenna, you can't hide or isolate targets effectively.
- Diffusion-FAS: This method consistently outperformed both, successfully isolating targets and hiding users even in very "noisy" (cluttered) environments.

Summary

This paper proposes a new way to control radio waves. Instead of just changing the signal, we change the physical shape of the antenna using a smart AI. This allows a device to either disappear from a radar's view by blending into the background noise or act as a filter to remove interference so a radar can see a specific target clearly. The AI acts like a master sculptor, instantly molding the antenna into the perfect shape to solve the problem, even when it only has partial information about the environment.

Technical Summary: Diffusion Fluid Antenna Systems for Resilient ISAC

1. Problem Statement

Current Integrated Sensing and Communication (ISAC) research predominantly focuses on the base station (BS) perspective, utilizing advanced waveform design and power allocation to support sensing and communication over shared resources. However, the "object-side" perspective remains underexplored. Two critical challenges arise from this gap:

Security and Stealth: There is a need to make objects difficult to detect for security reasons (e.g., hiding a target). Traditional methods like jamming are controversial as they pollute the entire radio environment, while purely AI-driven post-processing cannot alter the physical scattering properties of an object once illuminated.
Interference and Reliability: In close-proximity scenarios, a non-target object may generate dominant reflections that confuse the BS, creating interference that impairs the sensing reliability and localization accuracy of the intended target.

Existing solutions often rely on massive MIMO, which faces scalability limitations regarding power consumption and feedback overhead. Furthermore, optimizing ISAC solely in the power domain is insufficient for actively reshaping sensing signatures. The paper posits that Fluid Antenna Systems (FAS), which introduce a reconfigurable spatial degree of freedom (DoF) via port selection, offer a physical-layer solution to these challenges. However, controlling FAS to achieve specific sensing objectives (like stealth or interference nulling) is an NP-hard combinatorial problem, especially when the object has only partial observability of the channel state.

2. Methodology: Diffusion-FAS

The paper proposes Diffusion-FAS, a generative artificial intelligence (AI) framework that treats ISAC as a dynamic spatial selection problem. Instead of optimizing in the power domain, the system selects antenna ports to shape sensing signatures while maintaining communication objectives.

Core Architecture

Generative Policy: The system employs a Conditional Denoising Diffusion Probabilistic Model (DDPM). Rather than treating port selection as a discrete classification problem, the model operates in a continuous logit space.
Partial Observability Handling: Since the FAS-enabled node (User B) cannot observe the full channel state (only a sparse subset of ports is observed), the DDPM reconstructs the latent spatial correlation structure from these limited measurements.
Energy-Guided Sampling: To solve the NP-hard discrete mask selection problem, the framework uses a differentiable soft-top-M relaxation. During the reverse diffusion process, a customized sensing-energy function guides the trajectory toward optimal port configurations. This energy function is derived from the matched-filter overlap kernel and is differentiable, allowing gradients to steer the generation.

Operational Paradigms

The framework is designed to handle two distinct use cases by adjusting the energy guidance weights:

Generative Spatial Stealth (Use Case I): The FAS node actively selects ports to minimize its effective coupling magnitude, forcing its backscattered return to statistically emulate colored background clutter. The objective is to reduce the Kullback-Leibler (KL) divergence between the node's presence and the noise/clutter distribution, thereby suppressing detection probability.
Target Isolation / Cooperative Interference Shaping (Use Case II): The FAS node (User B) selects a sparse port mask to create deep spatial nulls in the "guard bins" surrounding a target-of-interest (User A). This actively rejects interference from the adjacent object, improving the target's localization accuracy under a fixed Constant False Alarm Rate (CFAR) threshold.

Training and Inference

Training: The model is trained offline via expert-imitation. A heuristic oracle exhaustively searches for high-quality logit vectors (minimizing the energy function) given full Channel State Information (CSI) to create a training dataset. The DDPM learns to predict the noise added to these expert solutions.
Inference: At runtime, the system generates multiple candidate masks from pure noise, guided by the learned prior and the partial observation context. The final hard mask is selected based on the minimum deterministic energy functional.

3. Key Contributions

The paper outlines four primary contributions:

Sensing-Centric Formulation: The authors formulate a sensing snapshot model over a whitened matched-filter bank with colored clutter. They introduce strict CFAR-calibrated metrics, specifically focusing on conditional localization accuracy ( $P(\text{correct}|\text{detected})$ ), to prove that FAS can protect the localization grid of adjacent targets.
Novel Operational Paradigms: The paper mathematically defines and solves two new paradigms: (i) clutter-likeness/stealth via intelligent port selection to evade sensing, and (ii) cooperative interference shaping to carve out spatial nulls in target guard bins.
Conditional DDPM Design: A conditional DDPM is designed to circumvent the NP-hard discrete mask selection problem. By diffusing over continuous logits and applying a differentiable energy metric derived from the matched-filter overlap kernel, the policy respects strict cardinality budgets while minimizing sensing interference.
Robust Numerical Evaluation: The paper presents results assessing the framework under varying hardware constraints (active ports, aperture size, measurement budget) and channel conditions (finite vs. rich scattering) without retraining the model for each scenario.

4. Numerical Results

The simulations, conducted with $K=200$ ports and varying observation budgets, demonstrate the following:

Target Isolation (Case 2):
- Port Count: As the number of active ports ( $M_{active}$ ) increases from 1 to 5, Diffusion-FAS exhibits a sharp, non-linear performance gain, rapidly closing the gap with the theoretical interference-free upper bound. Performance saturates at a compact subset ( $M_{active} \approx 5$ ), whereas random selection and static arrays fail to overcome proximity interference.
- Partial Observability: The system successfully reconstructs the spatial interference field using less than 10% of available ports ( $M_{obs} \approx 15-20$ ). Performance degrades under extreme sparsity ( $M_{obs} \le 10$ ) but recovers sharply as the observation window expands.
- Aperture Size: The framework effectively leverages spatial DoFs once the aperture exceeds $0.5\lambda$ . A compact $2\lambda$ footprint is sufficient to synthesize optimal spatial nulls.
- Clutter and Geometry: Diffusion-FAS maintains performance close to the interference-free bound even under severe clutter ( $\sigma_c$ up to 2.0) and when the interferer is adjacent to the target ( $\Delta \ge 1$ ). In contrast, baseline schemes (Random FAS, No FAS) suffer catastrophic performance drops in these scenarios.
- ROC Performance: The proposed method maintains a significant performance gap over baselines across various False Alarm Rates, ensuring that detections are spatially accurate.
Generative Stealth (Case 1):
- The framework achieves a two-order-of-magnitude reduction in the probability of detection ( $P_D$ ) compared to uninformed random selection.
- By navigating localized deep fades, the FAS node effectively "hides" its signature, maintaining a stealth floor significantly lower than static or fixed-subset baselines.

5. Significance and Claims

The paper claims that Diffusion-FAS offers a viable solution for secure and resilient ISAC by combining generative priors with reconfigurable apertures. Its significance lies in:

Shifting the Paradigm: Moving ISAC optimization from the power/waveform domain to the spatial selection domain, enabling active control over sensing signatures at the object level.
Solving Combinatorial Complexity: Providing a tractable, real-time solution to the NP-hard port selection problem using generative AI, which traditional deep learning struggles to adapt to dynamic constraints.
Physical-Layer Security: Demonstrating that objects can actively reduce their sensing visibility or mitigate interference without jamming, thereby preserving the radio environment while achieving security goals.
Resilience: Showing that the system remains effective under sparse measurements, varying hardware constraints, and harsh environmental conditions, offering a robust pathway to transcend traditional waveform-based ISAC limits.