💻 computer science

Secure Semantic Communication over Wiretap Channels: Rate-Distortion-Equivocation Tradeoff

This paper establishes the rate-distortion-equivocation tradeoff for secure semantic communication over wiretap channels by deriving single-letter bounds for a lossy joint source-channel coding model with correlated semantic and observed source components under two distinct encoder access scenarios.

Original authors: Denis Kozlov, Mahtab Mirmohseni, Rahim Tafazolli

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Denis Kozlov, Mahtab Mirmohseni, Rahim Tafazolli

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 sending a very important, complex message to a friend across a noisy, crowded room. But there's a catch: a spy is listening in on the conversation, and you need to make sure the spy understands nothing about the most sensitive parts of your message, while your friend still understands everything perfectly.

This paper is about finding the perfect balance for that scenario, but with a modern twist: Semantic Communication.

The Core Problem: The "Meaning" vs. The "Data"

In traditional communication, we treat all data as equal bits (0s and 1s). But in Semantic Communication, we realize that some parts of a message are just "observations" (like a blurry photo of a street), while other parts are the "meaning" or "semantics" (like the fact that a car is speeding toward a pedestrian).

The authors propose a system where:

The Semantic Part (The "Why"): This is the crucial, sensitive meaning. It must be kept secret from the spy.
The Observed Part (The "What"): This is the raw data or context. It might be okay if the spy sees this, or maybe it needs some protection, but it's less critical than the meaning.

The challenge is to send both parts over a "wiretap channel" (a channel where a spy is listening) while ensuring:

Your friend can reconstruct the message with enough quality (low distortion).
The spy learns as little as possible about the semantic meaning (high "equivocation" or confusion).

The Two Ways to Pack the Message

The paper explores two different ways the sender (the encoder) can prepare the message, using a creative analogy of packing a suitcase:

Case 1: The Blind Packer. The sender only sees the "Observed" part (the blurry photo) and has to guess the "Semantic" part (the meaning) based on how they usually relate. It's like trying to pack a suitcase for a trip to the beach just by looking at a map, without seeing the actual clothes.
Case 2: The All-Seeing Packer. The sender sees both the photo and the meaning. They know exactly what the trip is about and can pack perfectly.

The Finding: Unsurprisingly, Case 2 (seeing everything) allows for a much better performance. You can send the message faster or keep it more secure because you have more information to work with.

The Secret Sauce: A Three-Layer Security Strategy

To protect the message, the authors propose a clever coding scheme that acts like a Russian Nesting Doll or a Multi-Layered Cake:

Compression (The Squeeze): They first squeeze the data. By accepting a little bit of "loss" (blurry details), they naturally hide some information. If the data is already compressed, the spy has less to work with.
Encryption (The Lock): They use a shared secret key (like a password only the sender and receiver know) to scramble specific parts of the message.
Channel Hiding (The Noise): They use the physics of the channel itself. They send the message in a way that the friend's connection is clear, but the spy's connection is so noisy that the message looks like static.

The paper introduces a new "Superposition" technique. Imagine sending a letter where:

The Public Layer is the envelope (everyone sees it).
The Private Layer 1 is the letter inside, encrypted with a key.
The Private Layer 2 is a hidden compartment, protected by the channel noise itself.

This allows them to control exactly how much the spy knows about the "Observed" part versus the "Semantic" part independently.

The Trade-Off: The "Tightrope Walk"

The paper calculates the mathematical limits of this system, known as the Rate-Distortion-Equivocation Trade-off. Think of this as a tightrope walk between three competing goals:

Rate (Speed): How fast can you send the message?
Distortion (Quality): How blurry or imperfect can the message be at the receiver's end?
Equivocation (Confusion): How confused should the spy be?

The authors found that:

If you want perfect secrecy for the semantic part, you might have to send the message slower or accept a blurrier image.
If you have a secret key (password), you can send faster or keep better quality.
If the sender has direct access to the semantic meaning (Case 2), they can achieve much better results than if they only have the raw data (Case 1).

The Results: Gaussian and Binary Worlds

To prove their theory works in the real world, the authors tested their math on two specific types of data:

Gaussian (Smooth/Continuous): Like temperature readings or audio waves.
Binary (On/Off): Like simple yes/no switches or digital bits.

They created 3D maps (visualizations) showing exactly how much speed, quality, and security you can get for different settings. The results confirmed that Case 2 (seeing the meaning) always outperforms Case 1 (guessing the meaning), and that their new coding method gets very close to the theoretical best possible limit.

In Summary

This paper provides a mathematical blueprint for sending "meaningful" data securely. It shows that by treating the "meaning" and the "data" differently, and by using a mix of compression, encryption, and channel noise, we can protect sensitive semantic information much more efficiently than before. It proves that if you know what you are talking about (not just how it looks), you can send it more securely and efficiently.

Technical Summary: Secure Semantic Communication Over Wiretap Channels

Problem Statement
This paper investigates the fundamental limits of secure semantic-aware communication over wiretap channels. Unlike traditional communication models that treat the source as a single random variable, this work models a semantic source as two correlated random variables: a semantic component ( $S$ ) and an observed component ( $U$ ). The system employs lossy Joint Source-Channel Coding (JSCC) to transmit these correlated components over a memoryless wiretap channel to a legitimate receiver while preventing an eavesdropper from gaining information.

The core challenge addressed is the Rate-Distortion-Equivocation tradeoff. The system must satisfy:

Fidelity Constraints: Separate average distortion limits ( $D_s, D_u$ ) for the semantic and observed parts at the legitimate receiver.
Secrecy Constraints: Separate equivocation thresholds ( $\Delta_s, \Delta_u, \Delta_{su}$ ) for the semantic part, the observed part, and their joint version at the eavesdropper.
Encoder Access Scenarios: The paper evaluates two distinct encoding configurations:
- Case 1: The encoder has access only to the observed samples ( $U^k$ ) and must infer the semantic part ( $S^k$ ) via correlation.
- Case 2: The encoder has direct access to both the semantic ( $S^k$ ) and observed ( $U^k$ ) samples.

Methodology
The authors derive information-theoretic bounds for the achievable region of the system parameters (communication rate $r$ , secret key rate $R_k$ , distortions, and equivocations).

Converse Bounds: The paper establishes a single-letter converse bound. The derivation relies on bounding the rate-distortion functions (RDFs) and utilizing the properties of the wiretap channel. The converse explicitly incorporates standard RDFs and secrecy capacity terms, making it computationally tractable for common distributions. The proof involves bounding the mutual information between the source and the eavesdropper's observation, accounting for the shared secret key and the lossy compression.
Achievability Scheme: A novel four-layer stochastic superposition coding scheme is proposed. This scheme utilizes:
- Source Coding: Two auxiliary random variables ( $A$ and $B$ ) representing the semantic and observed parts, split into encrypted and wiretap-protected layers.
- Channel Coding: Two private layers ( $W_1, W_2$ ) and public layers ( $Q_1, Q_2$ ) to map source indices to channel codewords.
- Encryption: A shared secret key is used to encrypt specific indices of the source codebook, allowing for partial encryption.
- Separation of Control: The scheme allows for separate control over the equivocation of each source component, a feature not present in classical single-layer wiretap models.

Key Results

General Bounds: The paper provides single-letter converse and achievability bounds for the rate-distortion-equivocation region. The converse bound is expressed in terms of RDFs and secrecy capacities, while the achievability bound is derived via the proposed superposition coding scheme.
Specialized Models: The general results are specialized to:
- Gaussian Systems: Bivariate Gaussian sources over Gaussian wiretap channels. Closed-form solutions for marginal and joint RDFs are provided for both Case 1 and Case 2 encoders.
- Binary Systems: Bernoulli sources over Binary Symmetric Wiretap Channels. Closed-form RDFs and converse bounds are derived.
Numerical Evaluations:
- Performance Gain: Numerical results demonstrate that Case 2 encoding (direct access to semantic data) yields significant performance gains over Case 1, particularly in terms of achievable rate for a given distortion and secrecy level.
- Selective Secrecy: The results illustrate that enforcing secrecy only on the semantic component (while allowing the observed component to be less secure) can result in better overall performance compared to enforcing full secrecy on both components.
- Tightness: For the Gaussian system in the "no secrecy" regime, the inner (achievability) and outer (converse) bounds coincide. However, a gap remains between the bounds in scenarios requiring full semantic or joint secrecy.

Significance and Claims
The authors claim that this work generalizes several previously established problems, including:

Secure lossy JSCC of classic sources (where the source is a single variable).
Secure source coding of semantic sources without channel coding.
Secure JSCC with side information.

The paper emphasizes that the proposed model addresses the unique security challenges of semantic communication, where the "meaning" of data (semantic part) may require higher protection than the raw data (observed part). By characterizing the trade-off between rate, fidelity, and separate secrecy constraints, the work provides a theoretical foundation for designing secure semantic communication systems for next-generation networks (e.g., 6G). The authors modestly note that while the converse bound is explicit and easy to evaluate, the gap between the converse and achievability bounds in high-secrecy regimes suggests that the current achievability scheme may not be optimal, leaving room for future research into tighter schemes or the optimality of separation principles in this context.