Most of security issues in deep learning are based on human-imperceptible adversarial perturbation, which can fool image recognition models of deep learning and bring a serious security threats to many practical applications. However, how to construct a universal adversarial perturbation for images is still an open question. In this paper, we make fully use of a residual network to get a universal perturbation, and then utilize a loss network to perform the similarity measure of images to carry out the adversarial attack. Experiment results on the CIFAR-10 dataset show that our scheme can get an 89% attack success rate.

Multimedia data needs huge storage space, and application of multimedia data needs powerful capability of computing. Cloud computing can help owner of multimedia data to deal with it. But, multimedia data on cloud may reveal privacy of data owner, such as sex, hobbies, address, looks, and so on. Data owner can encrypt multimedia data for confidentiality before uploading it to cloud. However, encrypted multimedia data makes its utilization difficult. In this paper, we first discover pre-existing schemes have problems of huge storage space, security and low efficiency due to their inefficient and insecure algorithms. Then, we provide an effective and practical privacy-preserving scale-invariant feature transform (SIFT) scheme for encrypted image. It uses leveled homomorphic encryption based on our new encoding schemes, our new homomorphic comparison, division and derivative encryption. Our new secure SIFT scheme can realize higher computing efficiency, greatly reduce communication costs and interactive times between user and server, and perform correct feature key point detection, accurate feature point description and image matching. We evaluate security and efficiency of our new secure SIFT scheme, and compare our new secure SIFT scheme with other schemes in detail. The result shows that it is closest to the original SIFT algorithm.

With the advancement of Internet of Things (IoT), a large number of electronic devices are connected to the Internet. These connected electronic devices acquire and transmit information, and respond to any received actions. In the medical ecosystem, hospitals can implement medical diagnosis (MD) with medical sensors, especially for remote auxiliary MD. But, in this context, patients' privacy (PP) is of paramount importance, and confidentiality of medical data is crucial. Therefore, the main challenge ahead is how to realize remote auxiliary MD while protecting confidentiality of the medical data and ensuring PP. In this article, based on somewhat homomorphic encryption (SHE) scheme addressed by Junfeng Fan and Frederik Vercauteren (FV), we provide the first instance of a new efficient SHE scheme for homomorphic evaluation over single instruction multiple data (SIMD). We also implement a new set of efficient SIMD homomorphic comparison and division schemes. Based on these findings, we implement efficient privacy preserving and SIMD homomorphic surf and multiretina-image matching schemes. Offered functionalities include SIMD homomorphic feature point detection, multiretina-image matching, and lesion detection for the encrypted retinal image of diabetic retinopathy. Finally, we provide a proof-of-concept application implementation toward remote auxiliary diagnosis systems for diabetes in order to showcase the core security and privacy pillars of our solution. In the meantime, our IoT system designed with lattice-based cryptography preserves data confidentiality under quantum computation and quantum computers.