This thesis explores the use of variational methods in privacy protection and reinforcement learning. The focus is on the privacy of users' data during the training phase of a classification task, where users send their data to a service provider. To minimize privacy leakage, a highly distributed setting is considered, where each user trains their own model locally on their private data and only shares the trained model outputs with the server. We proposed two schemes based on variational autoencoder (VAE) and bounded information bottleneck autoencoder (BIB-AE) to learn a class-specific manifold representation. However, the assumption of having samples of only one class per user might not hold, and the limitation of distributed frameworks is in achieving high classification performance, resulting in a utility-privacy trade-off. To address these issues, we proposed a new framework based on contrastive learning to defend against reconstruction and attribute inference attacks. For defence against the reconstruction attack, the correlation of encoded features with the original data is minimized, while training an encoder with the supervised contrastive loss removes redundant information about the original image. In the attribute inference attack, an encoder trained with the supervised and private contrastive loss is proposed, while an obfuscator module is trained in an adversarial manner to preserve the privacy of private attributes while maintaining high classification performance.

Finally, we explored an application of decentralized classification based on variational autoencoders in the context of offline reinforcement learning.