Restricted Category Removal from Model Representations using Limited Data

Pratik Mazumder, Pravendra Singh, Mohammed Asad Karim

Abstract: Deep learning models are trained on multiple categories jointly to solve several real-world problems. However, there can be cases where some of the classes may become restricted in the future and need to be excluded after the model has already been trained on them (Class-level Privacy). It can be due to privacy, ethical or legal concerns. A naive solution is to simply train the model from scratch on the complete training data while leaving out the training samples from the restricted classes (FDR - full data retraining). But this can be a very time-consuming process. Further, this approach will not work well if we no longer have access to the complete training data and instead only have access to very few training data. The objective of this work is to remove the information about the restricted classes from the network representations of all layers using limited data without affecting the prediction power of the model for the remaining classes. Simply fine-tuning the model on the limited available training data for the remaining classes will not be able to sufficiently remove the restricted class information, and aggressive fine-tuning on the limited data may also lead to overfitting. We propose a novel solution to achieve this objective that is significantly faster ($\sim200\times$ on ImageNet) than the naive solution. Specifically, we propose a novel technique for identifying the model parameters that are mainly relevant to the restricted classes. We also propose a novel technique that uses the limited training data of the restricted classes to remove the restricted class information from these parameters and uses the limited training data of the remaining classes to reuse these parameters for the remaining classes. The model obtained through our approach behaves as if it was never trained on the restricted classes and performs similar to FDR (which needs the complete training data). We also propose several baseline approaches and compare our approach with them in order to demonstrate its efficacy.