The natural world is full of noise, but the brain’s capacity for information transmission is severely limited. Therefore, discarding irrelevant information contained in sensory inputs while retaining salient features that are related to the input label, is key to survival. What are the salient features? And what are the underlying feature selection mechanisms? It is thought that the brain may implement information bottlenecks, which aim to optimise the trade-off between compression and preservation of salient information. However, information bottlenecks are notoriously difficult to implement due to the intractability of mutual information in high dimensions (Paninski, 2003). In recent years, progress has been made by framing the estimation of mutual information as a minmax optimisation problem in an adversarial setting. Building on this, we propose a novel adversarial-inspired autoencoder framework, the objective of which is to compress data such that it can be accurately classified but cannot be fully reconstructed. We validated the efficiency of the framework on coloured MNIST digits and the CIFAR10 dataset, and our results show that it learns to discard irrelevant information, such as colour, while retaining salient information tied to the semantic label of the image. Additionally, the framework appears to perform figure-ground separation without explicit training, suggesting that it may mimic salient feature extraction in the early visual pathway. Our findings suggest the framework may find potential use in generating new, testable hypotheses about the salient features underlying noise-robust sensory information processing.