Compressing a Flow-Based Privacy Protection Model via a Novel Joint Distilling and Pruning Method

Abstract

While reversible anonymization techniques have been shown to be effective in protecting facial information, most of these methods rely on large reversible generative models that are only suitable for running on cloud servers, which poses a huge challenge to placing such privacy-preserving models on camera terminals. This work proposes a novel compression method that combines knowledge distillation and model pruning for training compact reversible generative models with an encoder-decoder pair. The proposed compression method exploits a novel collaborative knowledge distillation component to transfer the teacher model’s knowledge to the student model by pairing a teacher encoder (decoder) with a student decoder (encoder). Sub-sequently, a correlation-based regularized pruning component is adopted to further remove redundant parameters in the manually designed student model. This set of schemes enables us to install an anonymization system that builds privacy for sensitive personal data on embedded devices. We deploy and evaluate our compression privacy-preserving system on an NVIDIA Jetson TX2 device, which runs in real time and effectively anonymizes facial images to preserve privacy.