Convolutional neural networks (CNNs) have excellent and extensive applications in image recognition. With the continuous exploitation of data value and the proliferation of machine learning-as-a-service, convolutional neural network prediction schemes on privacy preservation have been introduced one after another, which makes much more attention focused on the privacy leakage and services offered to be efficient and light. Therefore, how to improve the convolutional neural prediction scheme on the premise of privacy preservation turns out to be an imperative research issue. In this paper, we propose a privacy-preserving convolutional neural network prediction scheme (PCP-LL) that supports low latency and lightweight users. The scheme starts from the perspective of lossless accuracy from underlying networks. First, we construct a secure activation function computing protocol (SActF) utilizing a commodity-based secure comparison protocol, which reduces the complexity and latency during the activation function computing under ciphertexts compared with common schemes. Second, to further support lightweight users, we introduce a secure output layer protocol (SOut) that enables users to obtain the prediction results without extra decryption after simple operations. Then, the scheme adopts the distributed two trapdoors public-key cryptosystem (DT-PKC) to achieve both data and model security, which well avoids security issues especially such as wiretapping by semi-honest participants commonly in secret sharing schemes. Finally, through relevant evaluations, the scheme not only achieves privacy preservation and low latency, but also supports lightweight users.

With the development of big data, deep learning has made remarkable progress on multi-view clustering. Multi-view fusion is a crucial technique for the model obtaining a common representation. However, existing literature adopts shallow fusion strategies, such as weighted-sum fusion and concatenating fusion, which fail to capture complex information from multiple views. In this paper, we propose a novel fusion technique, entitled contrastive fusion, which can extract consistent representations from multiple views and maintain the characteristic of view-specific representations. Specifically, we study multi-view alignment from an information bottleneck perspective and introduce an intermediate variable to align each view-specific representation. Furthermore, we leverage a single-view clustering method as a predictive task to ensure the contrastive fusion is working. We integrate all components into an unified framework called CONtrAstive fusion Network (CONAN). Experiment results on five multi-view datasets demonstrate that CONAN outperforms state-of-the-art methods. Our source code will be available soon at https://github.com/guanzhou-ke/conan.