⚡ electrical engineering

Secure Rate-Distortion-Perception: A Randomized Distributed Function Computation Approach for Realism

This paper characterizes the fundamental trade-offs between rate, distortion, and perception under strict security constraints for both noiseless and broadcast channels, demonstrating that randomized distributed coding with common randomness can simultaneously achieve strong secrecy, low distortion, and high perceptual quality.

Original authors: Gustaf Åhlgren, Onur Günlü

Published 2026-04-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gustaf Åhlgren, Onur Günlü

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 trying to send a high-definition photo of a sunset to a friend over a public Wi-Fi network. You want three things to happen:

The photo must look good: It shouldn't be a blurry mess (low distortion).
The photo must feel real: If you look at a thousand photos you send, they should all look like natural sunsets, not like weird, glitchy computer art (high perceptual quality).
The photo must be secret: A hacker sitting on the Wi-Fi shouldn't be able to figure out what the photo looks like just by listening to the data stream (security).

This paper is a mathematical blueprint for achieving all three of these goals at the same time. The authors, Gustaf Åhlgren and Onur Günlü, tackle a problem where usually, you have to sacrifice one thing to get another. They show you how to cheat the system using a clever trick called "Common Randomness."

Here is a breakdown of their ideas using simple analogies:

1. The Problem: The "Perfect Copy" vs. The "Real Feeling"

In old-school data compression, the goal was to send a file so the receiver could reconstruct the exact original bit-for-bit. But in modern AI and image compression, we don't need a pixel-perfect copy. We just need the feeling of the original.

Think of it like a painting.

Standard Compression: You try to send the exact brushstrokes. If you lose a few, the painting looks wrong.
Perception-Based Compression: You send a description of the style and colors. The receiver's computer (the decoder) uses its own "imagination" to paint a new sunset that looks just as beautiful as the original, even if the pixels are slightly different.

The Catch: When you send these "instructions" over a public channel, a hacker (Eve) might intercept them. If the instructions are too simple, the hacker can guess the image. If they are too complex to hide, you need a huge amount of bandwidth (slow internet).

2. The Secret Weapon: "Common Randomness"

The authors introduce a concept called Common Randomness. Imagine the sender and receiver share a secret, private notebook of random numbers that the hacker doesn't have.

The Analogy: Think of the data stream as a message in a bottle thrown into a public ocean.
- Without the notebook: The message inside is just the raw image data. If the hacker catches the bottle, they see the image.
- With the notebook: The sender writes the image data, but then scrambles it using a code from their shared notebook. The hacker catches the bottle, but it just looks like gibberish. The receiver catches the bottle, looks at their matching notebook, and unscrambles the code to reveal the beautiful sunset.

The paper proves that having this shared "notebook" (Common Randomness) allows you to send the message much faster (lower communication rate) while keeping it perfectly secure and looking realistic.

3. The Scenarios They Tested

The authors didn't just talk about theory; they looked at different "delivery methods":

The Perfect Pipe (Noiseless Channel): Imagine a fiber-optic cable where no data is lost. They figured out the exact mathematical limit of how fast you can send the image while keeping it secret and realistic.
The Noisy Radio (Broadcast Channel): Imagine sending a signal over the air where static interferes, and the signal goes to both your friend and the hacker. The hacker might get a slightly clearer signal than your friend (or vice versa). They found a way to ensure that even if the hacker gets a good signal, they still can't reconstruct the image.
The "Side Information" Bonus: Sometimes, the receiver already knows some things about the image (like, "I know this is a sunset, so I know the sky is blue"). The paper shows how to use this extra knowledge to send even less data.

4. The Big Discovery: Separation is Okay (Sometimes)

In engineering, there's a classic debate: Should you compress the data first and then send it (Separate), or do it all in one giant step (Joint)?

Usually, doing it all in one step is better.
However, the authors found that if you have unlimited shared randomness (a massive shared notebook), you can separate the tasks! You can compress the image first, then send it securely, and it works just as well as the complex, all-in-one method. This is great news because it makes building these systems much easier.

5. Why This Matters (The "So What?")

This isn't just math for math's sake. This is the future of AI-generated media.

Neural Image Compression: When your phone compresses a photo using AI, it's trying to save space while keeping the photo looking "real."
Privacy: If you are sending medical scans or private photos, you don't want the compression algorithm to leak information to hackers.

The paper shows that by using randomness (like a secret codebook), we can make these AI systems:

Faster: Send less data.
Safer: Hackers learn nothing.
Better: The images look more realistic.

Summary Metaphor

Imagine you are sending a secret recipe for a delicious cake to a friend.

The Old Way: You write the recipe on a postcard. It's fast, but anyone can read it. Or, you write it in a secret code, but the code is so long it takes forever to mail.
The New Way (This Paper): You and your friend share a magic deck of cards. You write the recipe, but you shuffle the words based on the cards you both hold. The postcard looks like random gibberish to a thief. Your friend uses their matching deck to unscramble it instantly.
The Result: The recipe arrives fast, it looks like a real recipe (not gibberish), and the thief learns nothing.

This paper provides the mathematical rules for how to build that "magic deck" system for the next generation of secure, high-quality AI communications.

1. Problem Statement

The paper addresses the fundamental trade-offs in Rate-Distortion-Perception (RDP) systems where security is a critical constraint.

Context: Modern applications (e.g., neural image compression, semantic communication) prioritize perceptual quality (realism) over bit-level fidelity. The goal is to reconstruct data $Y^n$ that not only has low distortion relative to the source $X^n$ but also follows a target probability distribution $Q_X$ similar to the source.
Security Challenge: When compressed data is transmitted over public channels, it is vulnerable to eavesdropping. The paper investigates Secure RDP, where the system must satisfy three simultaneous constraints:
1. Distortion: The expected distortion between source and reconstruction is bounded ( $D$ ).
2. Perception (Realism): The distribution of the reconstruction $P_{Y^n}$ must be close in Total Variation (TV) distance to the source distribution $Q_X^n$ .
3. Strong Secrecy: The mutual information between the reconstruction $Y^n$ and the eavesdropper's observation (channel output or transmitted index) must be negligible ( $\approx 0$ ).
Scenarios Analyzed:
- Noiseless communication channels.
- Noisy Broadcast Channels (BCs) with correlated noise components.
- Systems with Side Information (SI) available at the encoder, decoder, or both.
- Systems with unlimited Common Randomness (CR).

2. Methodology

The authors employ tools from Information Theory and Randomized Distributed Function Computation (RDFC).

Core Technique: The primary achievability proof method is Output Statistics of Random Binning (OSRB). This technique demonstrates that the induced probability distributions of a random binning scheme and a random coding scheme are asymptotically equal. By carefully designing the binning, the authors ensure the output statistics match the target distribution while satisfying secrecy constraints.
System Models:
- Noiseless Channel: The encoder sends an index $S$ to the decoder. An eavesdropper observes $S$ .
- Noisy BC: The encoder sends a channel input $\tilde{X}^m$ . The legitimate decoder receives $\tilde{Y}^m$ , and the eavesdropper receives $\tilde{Z}^m$ .
- Side Information: Models where side information $Z^n$ is correlated with the source $X^n$ and available at the encoder, decoder, or both.
Separation Analysis: The paper investigates whether Separate Source-Channel Coding is optimal for secure RDP. It compares joint coding schemes against concatenated source and channel codes, particularly under the assumption of unlimited common randomness.

3. Key Contributions

The paper makes several theoretical advancements:

Exact Secure RDP Region (Noiseless): Characterized the exact rate-distortion-perception region for noiseless channels under strong secrecy. The region is defined by the mutual information terms involving an auxiliary random variable $U$ .
Secure RDP over Noisy BCs:
- Derived an inner bound for the secure RDP region over noisy memoryless broadcast channels.
- Proved that for more-capable BCs (where the legitimate receiver's channel is strictly better than the eavesdropper's), this inner bound is tight, providing an exact rate region.
Side Information Extensions:
- Established the exact secure RDP region when side information is available at both the encoder and decoder.
- Derived an inner bound for the case where side information is available only at the decoder, and proved it is exact for a special case where the side information and output are jointly i.i.d.
Optimality of Separation: Demonstrated that for the secure RDP problem over a more-capable BC with unlimited common randomness, separate source-channel coding is optimal. This extends previous separation results to the secure, perceptual setting.
Role of Common Randomness: Provided examples showing that common randomness significantly reduces the required communication rate in secure RDP settings—a gain not observed in standard rate-distortion problems.

4. Key Results and Theorems

A. Noiseless Channels (Theorem 1)

The secure RDP region $\mathcal{R}$ is the union of tuples $(R, R_0, D)$ satisfying:

$R \geq I(U; X)$
$R_0 \geq I(U; Y)$
$D \geq E[d(X, Y)]$
Markov chain: $X - U - Y$
$P_Y \approx Q_X$ (in TV distance)
Insight: The common randomness rate $R_0$ required for realism and distortion is sufficient to satisfy the secrecy constraint.

B. Noisy Broadcast Channels (Theorem 2 & Corollary 1)

General Case: An inner bound is derived using auxiliary variables $W_1$ and $W_2$ to handle the channel noise and secrecy.
More-Capable BC: If $I(\tilde{X}; \tilde{Y}) \geq I(\tilde{X}; \tilde{Z})$ , the region is exact. The common randomness rate $R_0$ is reduced by the secure communication rate $I(\tilde{X}; \tilde{Y}) - I(\tilde{X}; \tilde{Z})$ .

C. Side Information (Theorems 3 & 4)

Encoder & Decoder SI: The rate $R$ depends on conditional mutual information $I(U; X|Z)$ . The secrecy constraint modifies the common randomness requirement to $R_0 \geq I(U; Y) - I(U; Z)$ .
Decoder-only SI: An inner bound is provided. If $Z^n$ and $Y^n$ are jointly i.i.d., the bound becomes exact.

D. Separation Optimality (Theorem 5)

For the noisy BC with unlimited common randomness, the achievable region using separate source and channel coding ( $\mathcal{R}_{N,UC,S}$ ) is equivalent to the joint coding region ( $\mathcal{R}_{N,UC,J}$ ). This implies that in the presence of unlimited CR, complex joint coding is not necessary to achieve the fundamental limits of secure RDP.

E. Numerical Examples

Binary Example: Shows that increasing common randomness allows for a substantial reduction in communication rate $R$ for a fixed distortion $D$ . For instance, a 40-87% increase in $R_0$ can reduce $R$ by 45-52% at low distortion levels.
Gaussian Example: Analyzes the trade-off with decoder-side information. It shows that higher correlation between the side information and the source reduces the required communication rate. In specific high-distortion regimes, zero communication is feasible if the side information is strong enough.

5. Significance and Impact

Foundational Limits: This work establishes the fundamental information-theoretic limits for trustworthy learned compression. It quantifies the cost of adding security (secrecy) and perceptual realism to data compression.
Practical Relevance: The results are directly applicable to Neural Image Compression and Semantic Communication, where generative models are used at the decoder to reconstruct high-quality images from compressed data. The paper proves that such systems can be made secure against eavesdroppers without sacrificing perceptual quality, provided sufficient common randomness is available.
Paradigm Shift: The finding that common randomness is a powerful resource for reducing communication rates in secure perceptual settings (unlike in standard R-D) suggests that future secure communication systems should leverage shared randomness (e.g., from previous frames or physical layer features) to optimize bandwidth usage.
Design Guidance: The proof of separation optimality under unlimited CR simplifies system design, suggesting that engineers can design robust source coders and channel coders independently for these specific secure scenarios.

In conclusion, the paper successfully integrates the concepts of perception (realism), security (secrecy), and efficiency (rate) into a unified framework, demonstrating that random binning-based coding can simultaneously achieve strong secrecy, low distortion, and high perceptual quality.