Adversarial Jamming for Autoencoder Distribution Matching

TL;DR

The paper addresses latent-space distribution matching for autoencoders by formulating a zero-sum minimax game between an adversarial jammer and a DeepJSCC autoencoder. Theoretical results imply the jammer's optimal noise is diagonal Gaussian, which is used as a regulariser to steer the latent distribution toward a diagonal Gaussian prior. Empirically, adversarial jamming yields distribution-matching performance comparable to VAEs and WAEs on CIFAR-10, CelebA, and MNIST, with gains tied to model capacity. This approach offers a principled, extensible route to latent-prior matching without explicit KL or MMD terms, potentially extendable to non-Gaussian priors.

Abstract

We propose the use of adversarial wireless jamming to regularise the latent space of an autoencoder to match a diagonal Gaussian distribution. We consider the minimisation of a mean squared error distortion, where a jammer attempts to disrupt the recovery of a Gaussian source encoded and transmitted over the adversarial channel. A straightforward consequence of existing theoretical results is the fact that the saddle point of a minimax game - involving such an encoder, its corresponding decoder, and an adversarial jammer - consists of diagonal Gaussian noise output by the jammer. We use this result as inspiration for a novel approach to distribution matching in the latent space, utilising jamming as an auxiliary objective to encourage the aggregated latent posterior to match a diagonal Gaussian distribution. Using this new technique, we achieve distribution matching comparable to standard variational autoencoders and to Wasserstein autoencoders. This approach can also be generalised to other latent distributions.

Figures (3)

• Figure 1: Adversarial jamming for prior distribution matching. The game between a compressor/jammer $f$ (compressing the images) and a JSCC autoencoder $(g, h)$ (for source $\mathbf{X} \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$) has a saddle point with $Z^* \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$, implicitly imposing this prior on the latent space.
• Figure 2: Reconstruction results for the CIFAR-10 dataset (top) and generation results for the CelebA dataset (bottom), for experiments using adversarial jamming, KL divergence (VAE), and maximum mean discrepancy (WAE) as the prior distribution latent regularisation term.
• Figure 3: The FID score of a generative model trained using adversarial jamming versus the hidden layer size of the DeepJSSC autoencoder. Increasing the capacity of the DeepJSCC autoencoder for learning improves the FID score of the generative model.