Deep Joint Source Channel Coding for Privacy-Aware End-to-End Image Transmission

TL;DR

This work addresses privacy-aware end-to-end image transmission over wireless channels when source and channel statistics are unknown and non-iid. It extends DeepJSCC by combining a privacy funnel-inspired objective with a variational bound on information leakage, formulating a data-driven minimax training framework against multiple adversaries (colluding or non-colluding) across AWGN and fading channels. A CNN/GDN-based encoder–decoder is trained alongside adversarial networks that predict private attributes, using a loss that blends reconstruction distortion (MSE and SSIM) with a leakage proxy via cross-entropy, achieving improved SSIM, reduced information leakage, and higher adversarial failure across CIFAR-10 and CelebA with extensive ablations. The approach yields a practical privacy-utility trade-off suitable for real-world 6G-like networks, demonstrating robustness to different channel models and offering a framework for privacy-aware end-to-end multimedia transmission.

Abstract

Deep neural network (DNN)-based joint source and channel coding is proposed for privacy-aware end-to-end image transmission against multiple eavesdroppers. Both scenarios of colluding and non-colluding eavesdroppers are considered. Unlike prior works that assume perfectly known and independent identically distributed (i.i.d.) source and channel statistics, the proposed scheme operates under unknown and non-i.i.d. conditions, making it more applicable to real-world scenarios. The goal is to transmit images with minimum distortion, while simultaneously preventing eavesdroppers from inferring certain private attributes of images. Simultaneously generalizing the ideas of privacy funnel and wiretap coding, a multi-objective optimization framework is expressed that characterizes the tradeoff between image reconstruction quality and information leakage to eavesdroppers, taking into account the structural similarity index (SSIM) for improving the perceptual quality of image reconstruction. Extensive experiments on the CIFAR-10 and CelebA, along with ablation studies, demonstrate significant performance improvements in terms of SSIM, adversarial accuracy, and the mutual information leakage compared to benchmarks. Experiments show that the proposed scheme restrains the adversarially-trained eavesdroppers from intercepting privatized data for both cases of eavesdropping a common secret, as well as the case in which eavesdroppers are interested in different secrets. Furthermore, useful insights on the privacy-utility trade-off are also provided.

Figures (18)

• Figure 1: Proposed learning-based system model for end-to-end image transmission against multiple eavesdroppers.
• Figure 2: Proposed deep neural networks at Alice (encoder) and Bob (decoder). The notation $w\times w \times f$ denotes a convolutional layer with $f$ filters (channel outputs) of spatial extent $w$. The notation $w\times w \times (f \times n_{\sf T})$ at the last convolutional layer of encoder indicates that we require $f$ filters for each of $n_{\sf T}$ antennas at the encoder output. Moreover, $s(\cdot)$ denotes the stride, which can be downsampling (at the encoder) or upsampling (at the decoder).
• Figure 3: Implemented DNN at each of the adversaries for extracting sensitive part of image files.
• Figure 4: Privacy-utility trade-off over CIFAR-10.
• Figure 7: Effect of $\alpha$ and $w$ on the privacy-utility trade-off