Privacy-Aware Communication Over the Wiretap Channel with Generative Networks

Ecenaz Erdemir, Pier Luigi Dragotti, Deniz Gündüz

Introduction

In secure communications, the "wiretap channel" is a fundamental model where a sender (Alice) transmits a message to a legitimate receiver (Bob) over a noisy channel, while a passive eavesdropper (Eve) intercepts the transmission through her own noisy channel. Traditional approaches focus on perfect secrecy, ensuring Eve learns nothing about the message. However, modern data sharing often requires a more nuanced "privacy-aware" approach. In this scenario, Alice aims to share a rich data source, such as an image, with Bob while specifically hiding certain sensitive latent attributes (e.g., identity, color, or thickness) from Eve.

This paper addresses the Privacy-Utility Trade-off (PUT) using a data-driven, end-to-end deep learning framework. Unlike traditional information-theoretic methods that require known data distributions, this approach utilizes Variational Autoencoders (VAEs) to handle complex, high-dimensional data like images.

System Architecture

The authors frame the problem as Joint Source-Channel Coding (JSCC) and implement it using a neural network architecture based on VAEs.

The proposed system model for privacy-aware communication, featuring an encoder-decoder pair for Alice and Bob and a classifier for the eavesdropper Eve.

The system consists of three main components:

  • Alice and Bob: Modeled as a VAE encoder-decoder pair. Alice encodes the source image into a discrete codeword for transmission over a Binary Symmetric Channel (BSC). Bob decodes the noisy received signal to reconstruct the original image.
  • Eve: Modeled as a neural network classifier that attempts to predict sensitive attributes from her version of the noisy codeword.
  • Optimization Strategy: The network is trained end-to-end to minimize reconstruction distortion for Bob, maximize mutual information (MI) between the source and Bob's signal for utility, and minimize MI between sensitive attributes and Eve's signal for privacy.

Experimental Setup

The framework was evaluated using a Colored MNIST dataset. The primary message is the digit identity, while the sensitive attributes to be hidden are the digit's color (Red, Green, Blue) and thickness (thin, medium, thick).

Key Results and Insights

The system successfully demonstrates a controllable Privacy-Utility Trade-off. By adjusting a privacy tuning parameter ($\lambda$), the model can prioritize either image reconstruction quality or attribute secrecy.

Visual results demonstrating the privacy-utility trade-off. At $\lambda=20$, Bob's reconstruction (middle row) remains accurate, while Eve's reconstruction (bottom row) loses sensitive color and thickness information.

Key Findings:

  • Effective Disentanglement: The network learns to disentangle sensitive features (color/thickness) from non-sensitive features (digit shape) through the channel bottleneck without needing explicit labels for the non-sensitive parts.
  • Sensitivity of Mutual Information: A small reduction in Mutual Information (approx. 0.06) resulted in a significant 20% drop in Eve's classification accuracy, validating MI as a robust privacy metric.
  • Emergent Smart Routing: In scenarios with parallel channels of varying noise levels, the model automatically routes sensitive data through channels where Bob has a clear signal but Eve experiences high noise, effectively "starving" the eavesdropper of useful information.

Limitations and Future Work

While effective, the approach relies on variational approximations of Mutual Information, as exact calculations are intractable for complex data. Future research could explore tighter bounds for MI optimization and scale the framework to higher-dimensional datasets, such as high-resolution facial images or video, over continuous wireless channels.

Summary updated

Abstract

We study privacy-aware communication over a wiretap channel using end-to-end learning. Alice wants to transmit a source signal to Bob over a binary symmetric channel, while passive eavesdropper Eve tries to infer some sensitive attribute of Alice’s source based on its overheard signal. Since we usually do not have access to true distributions, we propose a data-driven approach using variational autoencoder (VAE)-based joint source channel coding (JSCC). We show through simulations with the colored MNIST dataset that our approach provides high reconstruction quality at the receiver while confusing the eavesdropper about the latent sensitive attribute, which consists of the color and thickness of the digits. Finally, we consider a parallel-channel scenario, and show that our approach arranges the information transmission such that the channels with higher noise levels at the eavesdropper carry the sensitive information, while the non-sensitive information is transmitted over more vulnerable channels.
BibTeX
@inproceedings{erdemir2022privacy,
  author = {Ecenaz Erdemir and Pier Luigi Dragotti and Deniz Gündüz},
  title = {{Privacy-Aware Communication Over the Wiretap Channel with Generative Networks}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2022},
  doi = {10.1109/icassp43922.2022.9747068},
}
Figure