The vulnerability of deep neural networks to adversarial attacks has raised significant concerns about the security and reliability of artificial intelligence systems. Adversarial training is an effective approach to enhance adversarial robustness. To address the issue that existing methods adopt fixed adversarial sample generation strategies but neglect the importance of the adversarial sample generation phase for adversarial training, an adversarial training method was proposed based on adaptive attack strength. Firstly, the clean sample and the adversarial sample were input into the model to obtain the output. Then, the difference between the model outputs of the clean sample and the adversarial sample was calculated. Finally, the change of the difference compared with the previous moment was measured to automatically adjust the strength of the adversarial sample. Comprehensive experimental results on three benchmark datasets demonstrate that compared with the baseline method Adversarial Training with Projected Gradient Descent (PGD-AT), the proposed method improves the robust precision under AA (AutoAttack) attack by 1.92, 1.50 and 3.35 percentage points on three benchmark datasets, respectively, and the proposed method outperforms the state-of-the-art defense method Adversarial Training with Learnable Attack Strategy (LAS-AT) in terms of robustness and natural accuracy. Furthermore, from the perspective of data augmentation, the proposed method can effectively address the problem of diminishing augmentation effect during adversarial training.
With the popularity of cloud storage services and telemedicine platforms, more and more medical images are uploaded to the cloud. After being uploaded, the uploaded medical images may be leaked to unauthorized third parties, resulting in the disclosure of users’ personal privacy. Besides, if medical images are only uploaded to a single server for storage, they are vulnerable to attacks resulting in the loss of all data. To solve these problems, a medical image privacy protection algorithm based on thumbnail encryption and distributed storage was proposed. Firstly, by encrypting the thumbnail of the original medical image, the relevance of the medical images was preserved properly while achieving the encryption effect. Secondly, the double embedding method was adopted when hiding secret information, and data extraction and image recovery were performed separately to achieve Reversible Data Hiding (RDH) of the encrypted image. Finally, the distributed storage method based on polynomial shared matrix was used to generate n shares of the image and distribute them to n servers. Experimental results show that by using the encrypted thumbnail as carrier, the proposed algorithm exceeds the traditional security encryption methods on embedding rate. Even if the server is attacked, the receiver can recover the original image and private information as long as it receives no less than k shares. In the privacy protection of medical images, experiments were carried out from the aspects of anti-attack and image recovery, and the analysis results show that the proposed encryption algorithm has good performance and high security.
Focusing on the issues that the Reserving Room Before Encryption (RRBE) embedding algorithm requires a series of pre-processing work and Vacating Room After Encryption (VRAE) embedding algorithm has less embedding space, an algorithm of reversible data hiding in encrypted image based on multi-objective optimization was proposed to improve the embedding rate as well as reducing the algorithm process and workload. In this algorithm, two representative algorithms in RRBE and VRAE were combined and used in the same carrier, and performance evaluation indicators such as the amount of information embedded, distortion of direct decryption of image, extraction error rate, and computational complexity were formulated as the optimization sub-objectives. Then, the efficiency coefficient method was used to establish a model to solve the relative optimal solution of the application ratio of the two algorithms. Experimental results show that the proposed algorithm reduces the computational complexity of using RRBE algorithm alone, enables image processing users to flexibly allocate optimization objectives according to different needs in actual application scenarios, and at the same time obtains better image quality and a satisfactory amount of information embedding.
Rapid development of Location Based Service (LBS) and Augmented Reality (AR) technology lead to the hidden danger of user location privacy leakage. After analyzing the advantages and disadvantages of existing location privacy protection methods, a location privacy protection method was proposed based on location security. The zone security degree and the camouflage region were introduced into the method, and the zone security was defined as a metric that indicates whether a zone needs protection. The zone security degree of insecure zones (zones need to be protected) was set to 1 while that of secure zones (zones not need to be protected) was set to 0. And the location security degree was calculated by expanding zone security degree and recognition levels. Experimental results show that, compared with the method without introducing location security, this method can reduce average location error and enhance average security, therefore effectively protecting the user location privacy and increasing the service quality of LBS.
The Multidimensional Knapsack Problem (MKP) is a kind of typical multi-constraint combinatorial optimization problems. In order to solve this problem, a Greedy Binary Lion Swarm Optimization (GBLSO) algorithm was proposed. Firstly, with the help of binary code transform formula, the locations of lion individuals were discretized to obtain the binary lion swarm algorithm. Secondly, the inverse moving operator was introduced to update the location of lion king and redefine the locations of the lionesses and lion cubs. Thirdly, the greedy algorithm was fully utilized to make the solution feasible, so as to enhance the local search ability and speed up the convergence. Finally, Simulations on 10 typical MKP examples were carried out to compare GBLSO algorithm with Discrete binary Particle Swarm Optimization (DPSO) algorithm and Binary Bat Algorithm (BBA). The experimental results show that GBLSO algorithm is an effective new method for solving MKP and has good convergence efficiency, high optimization accuracy and good robustness in solving MKP.
In NarrowBand Internet of Things (NB-IoT) systems, the Internet of Things (IoT) terminals should decode Downlink Control Information (DCI) quickly to receive resource allocation and scheduling information of the data channel correctly. Therefore, a low complexity Narrowband Physical Downlink Control Channel (NPDCCH) blind detection algorithm using correlation detection was proposed for NPDCCH with search space size being greater than or equal to 32. By employing two correlation judgments on the data in a possible minimum repetition transmission unit of NPDCCH, the invalid data in searching space was removed to reduce the computation complexity. Then, the repetition periods with the valid data were combined and decoded to improve the blind detection performance. Finally, theoretical and simulation analysis of two correlation thresholds used in correlation detection were carried out. Results show that compared with conventional exhaustive blind detection algorithm, the decoding complexity of the proposed algorithm is reduced by at least 75% and the detection performance gain is increased by 2.5 dB to 3.5 dB. The proposed algorithm is more beneficial for engineering practice.
As the existing dynamic programming algorithm cannot quickly solve Discounted {0-1} Knapsack Problem (D{0-1}KP), based on the idea of dynamic programming and combined with New Greedy Repair Optimization Algorithm (NGROA) and core algorithm, a Greedy Core Acceleration Dynamic Programming (GCADP) algorithm was proposed with the acceleration of the problem solving by reducing the problem scale. Firstly, the incomplete item was obtained based on the greedy solution of the problem by NGROA. Then, the radius and range of fuzzy core interval were found by calculation. Finally, Basic Dynamic Programming (BDP) algorithm was used to solve the items in the fuzzy core interval and the items in the same item set. The experimental results show that GCADP algorithm is suitable for solving D{0-1}KP. Meanwhile, the average solution speed of GCADP improves by 76.24% and 75.07% respectively compared with that of BDP algorithm and FirEGA (First Elitist reservation strategy Genetic Algorithm).
Concerning that the parameter estimation in defogging algorithms based on image restoration is easy to cause the loss of scene information, a new defogging algorithm for single image was proposed. On the basis of the dark channel prior method, the atmospheric scattering model was analyzed and then the influence to dark channel image caused by fog distribution was summarized, which is the basis for adding fog to the outdoor images. The transmittance was estimated through the field depth relationship between the fog added reference image and the outdoor image to defogging. The algorithm used physical model and multiple images to complete the estimation of relevant parameters and had a better result in retaining scene information. The experimental results show that the proposed algorithm is more effective than the comparison algorithms, and its processing speed is also improved significantly.
Question classification is one of the tasks in question answering system. Since questions often have rare words and colloquial expressions, especially in the application of voice interaction, the traditional text classifications perform poorly in short question classification. Thus a short question classification algorithm was proposed, which was based on semantic extensions and used the search engine to extend knowledge for short questions, the question's category was got by selecting features with the topic model and calculating the word similarity. The experimental results show that the proposed method can get F-measure value of 0.713 in a set of 1365 real problems, which is higher than that of Support Vector Machine (SVM), K-Nearest Neighbor (KNN) algorithm and maximum entropy algorithm. Therefore, the accuracy of the question classification can be improved by above method in question answering system.