Latent Space Stochastic Perturbation Schedule via Cosine Annealing Scheduler for Privacy-Preserving Variational Autoencoders

In this study, we leveraged cosine annealing scheduling mechanism to adaptively inject controlled perturbation into the latent space of VAE across training epochs to modify the standard reparameterization trick for better privacy protection and higher reconstruction fidelity. We used the Kraskov-Stogbauer-Grassberger (KSG) estimator to assess the impact of our technique on privacy protection while its impact on reconstruction fidelity was assesed with the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) metrics. Experiments on MNIST dataset indicated that our proposed LSSPS-VAE model outperformed the baseline model across all the performance metrics. The results suggest that latent space stochastic perturbation through cosine-annealed scheduler can be an effective technique for achieving better privacy protection and higher reconstruction fidelity in variational autoencoders.