Despite inherent ill-definition, anomaly detection is research endeavour of great interest within machine learning visual scene understanding alike. Most commonly, anomaly is considered as the detection of outliers within a data distribution based on some measure of normality. most significant challenge in real-world anomaly detection is that available data is highly imbalanced towards (i. e. non-anomalous) and contains at most a sub-set all possible anomalous samples - hence limiting the use of -established supervised learning methods. By contrast, we an unsupervised anomaly detection model, trained only the normal (non-anomalous, plentiful) samples in order to the normality distribution of the domain, and hence detect based on deviation from this model. Our proposed employs an encoder-decoder convolutional neural network with skip connections to thoroughly capture the multiscale distribution of the normal data distribution in image space. , utilizing an adversarial training scheme for this architecture provides superior reconstruction both within space and a lower-dimensional embedding vector space. Minimizing the reconstruction error metric within both image and hidden vector spaces during training aids the to learn the distribution of normality as required. Higher metrics during subsequent test and deployment thus indicative of a deviation from this normal distribution, indicative of an anomaly. Experimentation over established anomaly detection benchmarks and challenging real-world, within the context of X-ray security screening, shows unique promise of such a proposed approach.