Pretrained deep models hold their learnt knowledge in the form of model\nparameters. These parameters act as "memory" for the trained models and help\nthem generalize well on unseen data. However, in absence of training data, the\nutility of a trained model is merely limited to either inference or better\ninitialization towards a target task. In this paper, we go further and extract\nsynthetic data by leveraging the learnt model parameters. We dub them "Data\nImpressions", which act as proxy to the training data and can be used to\nrealize a variety of tasks. These are useful in scenarios where only the\npretrained models are available and the training data is not shared (e.g., due\nto privacy or sensitivity concerns). We show the applicability of data\nimpressions in solving several computer vision tasks such as unsupervised\ndomain adaptation, continual learning as well as knowledge distillation. We\nalso study the adversarial robustness of lightweight models trained via\nknowledge distillation using these data impressions. Further, we demonstrate\nthe efficacy of data impressions in generating data-free Universal Adversarial\nPerturbations (UAPs) with better fooling rates. Extensive experiments performed\non benchmark datasets demonstrate competitive performance achieved using data\nimpressions in absence of original training data.\n