Privacy-Preserving Action Recognition (PPAR) aims to transform raw videos into anonymous ones to prevent privacy leakage while maintaining action clues, which is an increasingly important problem in intelligent vision applications. Despite recent efforts in this task, it is still challenging to deal with novel privacy attributes and novel privacy attack models that are unavailable during the training phase. In this paper, from the perspective of meta-learning (learning to learn), we propose a novel Meta Privacy-Preserving Action Recognition (MPPAR) framework to improve both generalization abilities above (i.e., generalize to and) in a unified manner. Concretely, we simulate train/test task shifts by constructing disjoint support/query sets w.r.t.

Impact of constructing attribute-different (and model-different) support set and query set. As shown in Tab. 1, our framework ( Attribute/Model-Different Set Construction) achieves better performance than this variant, which demonstrates the Here we test two variants. As shown in Tab. 3, our Since our approach does not change the anonymization model's Face and Gender are novel attributes that are unknown during training. In our proposed framework, we first train the model using the support set (i.e., virtual training), and We explain the two reasons in more details below. The list of training/testing attributes and attack models.

Privacy-Preserving Action Recognition (PPAR) aims to transform raw videos into anonymous ones to prevent privacy leakage while maintaining action clues, which is an increasingly important problem in intelligent vision applications. Despite recent efforts in this task, it is still challenging to deal with novel privacy attributes and novel privacy attack models that are unavailable during the training phase. In this paper, from the perspective of meta-learning (learning to learn), we propose a novel Meta Privacy-Preserving Action Recognition (MPPAR) framework to improve both generalization abilities above (i.e., generalize to novel privacy attributes and novel privacy attack models) in a unified manner. Concretely, we simulate train/test task shifts by constructing disjoint support/query sets w.r.t. Then, a virtual training and testing scheme is applied based on support/query sets to provide feedback to optimize the model's learning toward better generalization.