Abstract:
Model parameters store pretrained deep models’ knowledge. These parameters give trained models “memory” to generalize well on unseen data. In the absence of training data, a trained model is only useful for inference or task initialization. This paper uses learned model parameters to extract synthetic data. Data Impressions, which represent training data, can be used for a variety of tasks. These are useful when only pretrained models are available and training data is not shared (e.g., privacy or sensitivity concerns). Data impressions solve unsupervised domain adaptation, continual learning, and knowledge distillation in computer vision. We examine the adversarial robustness of lightweight models trained via knowledge distillation using these data impressions. We show that data impressions can generate data-free Universal Adversarial Perturbations (UAPs) with better fooling rates. Extensive experiments on benchmark datasets show competitive performance using data impressions without training data.
Note: Please discuss with our team before submitting this abstract to the college. This Abstract or Synopsis varies based on student project requirements.
Did you like this final year project?
To download this project Code with thesis report and project training... Click Here
