Tensor-Based Privacy Protection Scheme With Multifeature Fusion for Facial Recognition
提出一种基于多特征融合张量的隐私保护方案，通过结合深度特征与手工特征生成扰动图像，在保持视觉相似性的同时提高人脸识别误判率，实验显示误判率提升16%。
As accurate face recognition (FR) models based on deep learning can be easily trained using face images from various social media platforms, this phenomenon has raised ever-increasing concerns regarding user privacy. To address this issue, we investigate a privacy protection scheme based on multifeature fusion tensor (PPS-MFFT). Different from previous studies using a single feature or simple combination of several features, for every face image, PPS-MFFT first constructs a multifeature fusion tensor through hierarchically exploiting the correlations and complementarity between deep-learning features and those handcrafted features for stronger robustness and transferability. Further, on the basis of such tensors, the target images are reasonably chosen to enhance the camouflage effects while maintaining the visual similarities for final perturbed images, which are generated by means of developing a new optimization model for better tradeoff between effectiveness and practicability. Finally, the measurement results validate that both higher protective efficacy (e.g., 16% more in misidentifying the original face images) and acceptable visual effects can be obtained by PPS-MFFT when compared to the existing methods, and thus demonstrate the generality and applicability of our scheme.