Meta-VSA: No-Reference Visual Security Assessment for Perceptual Encrypted Images

In the paper, a new Meta-VSA method is proposed to guide visual security assessment (VSA) for perceptual encrypted images. Specifically, the meta-learning is first applied to obtain prior knowledge and get a prior model, then the target visual security assessment task is fine-tuned to get the final scores. During the process, the images are randomly selected and introduced into the learning network with naturalness description. The experiments on two publicly available perceptual encryption image databases showed that our proposed is superior to the mainstream reference-free quality evaluation method and it outperforms the existing state-of-the-art methods in terms of accuracy and stability.