Crimes nowadays pose unique issues to security and legal institutions and requires smart approaches to different types of peculiar behavior within. This paper proposes a deep learning autoencodes framework to analyze and recognize unusual activities in the FBI’s crime dataset. Utilizing the autoencoder model’s architecture consisting of input, compression, and output layers, the Adam optimizer is used with a Mean Squared Error loss function for training, validating with twenty percent of the data. A reconstruction error is calculated and subsequently, a threshold of the 95th percentile of the average MSE is set to flag anomalies. Findings prove that the model outperforms all comparative methodologies, achieving 98% accuracy and a 97% precision F1 score. In addition, the model was shown to have an AUC on ROC curve of 98.2% which confirms the model’s ability to accurately classify normal and abnormal samples. This study illustrates the capability of multi-dimensional autoencoders to analyze and process complex crime data which can greatly aid security agencies in premeditative and reactive responses to crime. Further research will focus on attention-based hybrid models along with system for real-time responsive tracing of volatile hyperdynamics.

The short-and the long-term information in speech signal are useful for speech enhancement, especially if the speech signal is corrupted by both stationary and non-stationary noises. This paper proposes a new approach to provide long-term speech input for a deep denoising autoencoder by reducing the number of frequency sub-bands of the input data. This paper also proposes a two phase speech enhancement approach. The first phase performs short-term speech enhancement by using a deep denoising autoencoder. In the second phase, long-term speech enhancement denoising autoencoder is applied on the output of short-term enhanced speech data. The proposed models were evaluated on the Aurora-2 Speech recognition corpus and our results show significant improvements of 0. 3 in PESQ score at lower SNR values. The proposed models were evaluated on the recognition task where the proposed method results in 4% reduction in word error rate for the multi-condition training when compared to the baseline MFCC front-end.

Botnet is a group of hosts infected with the same malicious code and managed by an attacker or Botmaster through one or more command and control (C&C) servers. The new generation of Botnets generates C&C domain name server’ s list dynamically. This dynamic list created by a domain generation algorithm helps an attacker to periodically change its C&C servers and prevent their addresses from being blacklisted. Each infected host generates a large number of domain names using a predefined algorithm and attempts to map them to their corresponding addresses by sending queries to the domain server. In this paper, the deep autoencoder neural network is used to identify domains without any knowledge of their generating algorithm, and the performance of the proposed method is compared with the performance of machine learning algorithms. Initially, a new dataset is created by combining a data set with normal domains and two datasets containing malicious and abnormal domains and both manual and automated methods are used to extract the features of the new dataset. deep autoencoder neural network is applied to new and pre-processed datasets and the results are compared with machine learning algorithms. Based on the obtained results, it is possible to identify the malicious domains generated by domain generating algorithms using the deep autoencoder neural network with a higher speed and an accuracy rate larger than 98. 61%.

Text clustering is a method for separating specific information from textual data and can even classify text according to topic and sentiment, which has drawn much interest in recent years. deep clustering methods are especially important among clustering techniques because of their high accuracy. These methods include two main components: dimensionality reduction and clustering. Many earlier efforts have employed autoencoder for dimension reduction; however, they are unable to lower dimensions based on manifold structures, and samples that are like one another are not necessarily placed next to one another in the low dimensional. In the paper, we develop a deep Text Clustering method based on a local Manifold in the autoencoder layer (DCTMA) that employs multiple similarity matrices to obtain manifold information, such that this final similarity matrix is obtained from the average of these matrices. The obtained matrix is added to the bottleneck representation layer in the autoencoder. The DCTMA's main goal is to generate similar representations for samples belonging to the same cluster; after dimensionality reduction is achieved with high accuracy, clusters are detected using an end-to-end deep clustering. Experimental results demonstrate that the suggested method performs surprisingly well in comparison to current state-of-the-art methods in text datasets.

Recently, a number of Extreme Learning Machine (ELM) based training algorithms have been introduced for training deep neural network structures. ELM based Auto-Encoder (ELM-AE) is one such algorithm that has been used for making multilayer structures and tuning parameters of each layer. In a simple ELM-AE training algorithm, the weights of the first layer are initialized randomly. This issue is a leading factor in producing reconstruction error. The frequent use of ELM-AE in deep network layers results in propagating such errors through deep structures and in decreasing performance as a consequent. In this paper, we introduce a multilayer structure and a new learning algorithm to train it that prevents error propagation. In order to boost the performance of the model, the parameters in the first layer are initialized by a novel type of ELM-AE called Repeated-AE (RAE) rather than by a random selection method. This RAE-based technique determines the parameters in the first layer far better than do the other ELM-AE existed methods. Next, a single hidden layer ELM is applied for handling the classification task. Experimental results for data classification show that the proposed method outperforms some other methods in terms of the average accuracy over all datasets by amounts of 4%, 26%, 17% and 31%. Eventually, so as to verify the performance of the proposed multilayer ELM-AE in application, we used this model to reconstruct images. The reconstructed images obtained by our approach appeared visually a lot better compared to those obtained by the other methods do.

Demand for personalized recommendation systems elevated recently by e-commerce, news portals etc., to grab the customer interest on the sites. Collaborative filtering proves to be powerful technique but it always suffers from data sparsity, cold-start and robustness issues. These issues have been tackled by some approaches resulting in higher accuracy. Few of them take user profiles, item attributes and rating time as the side information along with ratings to give interpretative personalized recommendations. These type of approaches tries to find which factors mainly impacted the user to rate an item. Another approach extends the single-criteria ratings of collaborative filtering to multi-criteria ratings. Our approach exploits non-linear interpretative recommendations by exploring Multi-criteria ratings by combination of autoencoders with dropout layer and firefly algorithm optimized weights for deep neural networks. Our approach solves data sparsity, scalability issues and fetch accurate recommendations. Experimental evaluations have been done using Yahoo! Movie and MovieLens datasets. Our approach outperforms in robustness and accuracy with respect to previous research works.

Background. Brain computer interface (BCI) systems by extracting knowledge from brain signals provide a connection channel to the outside world for disabled people, without physiological interfaces. Event-related potentials (ERPs) are a specific type of electroencephalography signals and P300 is one of the most important ERP components. The critical part of P300-based BCI systems is classification step. In this research, an approach is proposed for P300 classification based on novel machine learning methods using convolutional neural networks (CNN) and autoencoder networks. Methods. In the pre-processing step, channel selection, data augmentation (by ADASYN method), filtering and base-line drift were done. Then, in the classification step, four different CNN classifiers including CNN1D, CNN2D, CNN1D_autoencoder, and CNN2D-autoencoder were used for P300 classification. Results. After implementation and tuning the networks, 92% as a best accuracy was achieved by CNN2D_autoencoder. This result was achieved with a considerable tradeoff between complexity and stability. Conclusion. The acquired results emphasize the ability of the deep learning methods in P300 classification and approve the advantage of using them in BCI systems. Furthermore, autoencoder versions of CNN networks are more stable and have a faster convergence. Meanwhile, ADASYN is a suitable method for augmentation of P300 data and even ERPs by sustaining the premier feature space without copying data. Practical Implications. Our results can increase the accuracy of P300 detection and simultaneously reduce the volume of data using the proposed model. Consequently, they can improve character recognition in P300-speller systems generally used by amyotrophic lateral sclerosis (ALS) patients.

Anomaly detection means detecting samples that are different from the normal samples in the dataset. One of the great challenges in this area is finding labeled data, especially for the abnormal categories. In this paper, we propose a method that uses normal data to detect anomalies. This method is based on established neural networks which are called automated encoder and are considered in deep learning studies. An automated encoder reproduces its input as output and reconstruction deviation to rate anomalies. We have used LSTM blocks to construct encoder instead of using ordinary neurons. In fact, these blocks are a category of recurring neural networks that are specialized in discovering and fetching time and proximity dependencies. The result of employing an automated encoder using LSTM blocks to detect point anomalies shows that this approach has been promising and successful in extracting the normal data’ s internal model and also detecting anomalous data. The AUC factor of the model, in almost all cases, is better than the AUC of an ordinary automated encoder and One Class Support Vector Machine (OC-SVM).

Graph embedding is the procedure of transforming a graph into a low-dimensional, informative representation. The majority of existing graph embedding techniques have given less consideration to the embedding distribution of the latent codes and more attention to the graph’s structure. Recently, Variational Graph autoencoders (VGAEs) have demonstrated good performance by learning smooth representations from unlabeled training samples. On the other hand, in regular VGAEs, the prior distribution over latent variables is generally a single Gaussian distribution. However, complex data distributions cannot be well-modelled under the assumption of a single Gaussian distribution. This choice of prior distribution is important because each dimension of a multivariate Gaussian can learn a separate continuous latent feature, which can result more structured and disentangled representation. In this paper, we employ the Gaussian Mixture Model (GMM) as the prior distribution in a Variational Graph autoencoder (GMM-VGAE) framework for node classification in graphs. In this framework, GMM effectively discovers the inherent complex data distribution, and graph convolutional networks (GCNs) exploit the structure of the nodes of a graph to learn more informative representations. The proposed model incorporates several Graph Convolutional Networks (GCNs): one to map the input feature vector to the latent representation utilized for classification, another to generate the parameters of the latent distribution for learning from unlabeled data, and finally, an additional GCN is employed for reconstructing the input and delivering the reconstruction loss. Through extensive experiments on well-known Citations, Co-authorship, and Social network graphs, GMM-VGAE’s superiority over state-of-the-art methods is demonstrated.

Distributed Denial of Service (DDoS) attacks are among the primary concerns in internet security today. Machine learning can be exploited to detect such attacks. In this paper, a multi-layer perceptron model is proposed and implemented using deep machine learning to distinguish between malicious and normal traffic based on their behavioral patterns. The proposed model is trained and tested using the CICDDoS2019 dataset. To remove irrelevant and redundant data from the dataset and increase learning accuracy, feature selection is used to select and extract the most effective features that allow us to detect these attacks. Moreover, we use the grid search algorithm to acquire optimum values of the model’s hyperparameters among the parameters’ space. In addition, the sensitivity of accuracy of the model to variations of an input parameter is analyzed. Finally, the effectiveness of the presented model is validated in comparison with some state-of-the-art works.