Automated Fraudulent Phone Call Recognition through Deep Learning

Investigating the learning approaches of students in nursing education

Telephone spam has become an increasingly prevalent problem in many countries all over the world. For example, the US Federal Trade Commission's (FTC) National Do Not Call Registry's number of cumulative complaints of spam/scam calls reached 30.9 million submissions in 2016. Naturally, telephone carriers can play an important role in the fight against spam. However, due to the extremely large volume of calls that transit across large carrier networks, it is challenging to mine their vast amounts of call detail records (CDRs) to accurately detect and block spam phone calls. This is because CDRs only contain high-level metadata (e.g., source and destination numbers, call start time, call duration, etc.) related to each phone calls. In addition, ground truth about both benign and spam-related phone numbers is often very scarce (only a tiny fraction of all phone numbers can be labeled). More importantly, telephone carriers are extremely sensitive to false positives, as they need to avoid blocking any non-spam calls, making the detection of spam-related numbers even more challenging. In this paper, we present a novel detection system that aims to discover telephone numbers involved in spam campaigns. Given a small seed of known spam phone numbers, our system uses a combination of unsupervised and supervised machine learning methods to mine new, previously unknown spam numbers from large datasets of call detail records (CDRs). Our objective is not to detect all possible spam phone calls crossing a carrier's network, but rather to expand the list of known spam numbers while aiming for zero false positives, so that the newly discovered numbers may be added to a phone blacklist, for example. To evaluate our system, we have conducted experiments over a large dataset of real-world CDRs provided by a leading telephony provider in China, while tuning the system to produce no false positives. The experimental results show that our system is able to greatly expand on the initial seed of known spam numbers by up to about 250%.

Missing data in longitudinal cognitive assessments may occur due to logistical (e.g., scheduling conflicts), personal (e.g., reduced motivation), and health-related issues (e.g., deteriorated physical or cognitive functions, anxiety, etc.). This issue poses significant challenges for clinical research and patient monitoring. Deep learning-based imputation approaches, which are agonistic to assumptions of data distribution and covariance structure, outperform traditional methods in imputing longitudinal missing data in other domains. We aim to evaluate their effectiveness in the domain of Alzheimer's Disease (AD), an area that has not been extensively explored. We developed a deep learning imputation approach using the self-attention-based imputation for time series (SAITS) model and de-identified data accessed on the AD Data Initiative's AD workbench. SAITS imputes missing values in multivariate time series by leveraging self-attention mechanisms to capture both temporal dependencies and feature correlations. It is trained by jointly optimizing imputation and reconstruction tasks on training data. We trained and evaluated SAITS using the Mini-Mental State Examination (MMSE) and Alzheimer's Disease Assessment Scale-Cognitive (ADAS-cog) test scores from GERAS-EU and GERAS-US studies, collected every 6 months over 3 years (7 visits in total). We compared SAITS' performance with (1) two state-of-the-art deep learning approaches: iTransformer and BRITS and (2) two conventional approaches: mean (averaging data from other visits of the same participant) and last observation carried forward (LOCF). We randomly split the full dataset into 80:20 training:test datasets. Deep learning models were optimized on the training set using 5-fold cross-validation and evaluated on the test set with 10% of values randomly masked/removed. Model performance was evaluated using mean absolute error (MAE) across 100 trials. Data from 1336 GERAS-EU participants (age: 72.0±11.6, 61.3% female) and 563 GERAS-US participants (age: 71.0±7.8, 52.7% female) were analyzed. SAITS performed best in most imputation tasks (Tables 1 and 2), achieving 1.636 (GERAS-EU) and 1.628 (GERAS-US) MAEs when imputing ADAS-cog total scores and 0.567 (GERAS-EU) and 1.178 (GERAS-US) MAEs when imputing MMSE total scores. Deep learning techniques demonstrates great potential in imputing missed longitudinal cognitive data in AD. Future work will evaluate their effects on down-stream tasks such as predicting cognitive decline.

Recently, neural network-based deep learning methods have been popularly applied to computer vision, speech signal processing and other pattern recognition areas. Remarkable success has been demonstrated by using the deep learning approaches. The purpose of this article is to provide a comprehensive survey for the neural network-based deep learning approaches on acoustic event detection. Different deep learning-based acoustic event detection approaches are investigated with an emphasis on both strongly labeled and weakly labeled acoustic event detection systems. This paper also discusses how deep learning methods benefit the acoustic event detection task and the potential issues that need to be addressed for prospective real-world scenarios.

Water quality monitoring is essential for managing water resources and ensuring human and environmental health. However, obtaining reliable data can be challenging and costly, especially in complex systems such as estuaries. To address this problem, we propose a novel deep learning-based approach that uses limited available data to accurately estimate and reconstruct critical water quality variables, such as electrical conductivity, dissolved oxygen, and turbidity. Our approach included two tasks, numerical modeling and historical reconstruction, and was applied to the Seine River in the Normandy region of France at four quality stations. In the first task, we evaluated four deep learning approaches (GRU, BiLSTM, BiLSTM-Attention, and CNN-BiLSTM-Attention) to numerically simulate each variable for each station under different input data selection scenarios. We found that incorporating the quality data with the water level data collected at the various stations into the input data improved the accuracy of the water quality data simulation. Combining water levels from multiple stations reliably reproduced electrical conductivity, especially at stations near the sea where tidal fluctuations control saltwater intrusion in the area. While each model had its strengths, the CNN-BiLSTM-Attention model performed best in complex tasks with dissimilar input trends, and the GRU model outperformed other models in simple monitoring tasks with similar input-target trends. The second task involved automatically searching the optimal configurations for completing the missing historical data in sequential order using the modeling task results. The electrical conductivity data were filled before the dissolved oxygen data, which were in turn more reliable than the turbidity simulation. The deep learning models accurately reconstructed 15 years of water quality data using only six and a half years of modeling data. Overall, this research demonstrates the potential of deep learning approaches with their limitations and discusses the best configurations to improve water quality monitoring and reconstruction.

This paper explores the effectiveness of machine learning and deep learning approaches in detecting deepfake images. With the increasing sophistication and prevalence of deepfake technology, it is becoming increasingly important to develop reliable methods for detecting manipulated images. The paper outlines the steps involved in creating a deepfake image detection model, including data preprocessing, dataset splitting, hyperparameter tuning, and the use of deep learning and machine learning techniques. Comprehensive review of the literature on deepfake detection, highlighting the challenges and limitations of existing approaches. They then propose a novel deep learning-based approach that leverages the power of convolutional neural networks (CNNs) to detect deepfake images. The proposed approach involves training a CNN on a large dataset of real and manipulated images, and using transfer learning to fine-tune the model on a smaller dataset of deepfake images. Here we try to evaluate the performance of their approach on a benchmark dataset of deepfake images, and compare it to several state-of-the-art deepfake detection methods. The results show that the proposed approach outperforms existing methods in terms of accuracy, precision, and recall. This paper provides valuable insights into the use of machine learning and deep learning approaches for deepfake image detection, and presents a promising new approach that could help to mitigate the risks of digital disinformation.

Accurately and early diagnosis of melanoma is one of the challenging tasks due to its unique characteristics and different shapes of skin lesions. So, in order to solve this issue, the current study examines various deep learning-based approaches and provide an effective approach for classifying dermoscopic images into two categories of skin lesions. This research focus on skin cancer images and provides solution using deep learning approaches. This research investigates three approaches for classifying skin cancer images. (1) Utilizing three fine-tuned pre-trained networks (VGG19, ResNet18, and MobileNet_V2) as classifiers. (2) Employing three pre-trained networks (ResNet-18, VGG19, and MobileNet v2) as feature extractors in conjunction with four machine learning classifiers (SVM, DT, Naïve Bayes, and KNN). (3) Utilizing a combination of the aforementioned pre-trained networks as feature extractors in conjunction with same machine learning classifiers. All these algorithms are trained using segmented images which are achieved by using the active contour approach. Prior to segmentation, preprocessing step is performed which involves scaling, denoising, and enhancing the image. Experimental performance is measured on the ISIC 2018 dataset which contains 3300 images of skin disease including benign and malignant type cancer images. 80% of the images from the ISIC 2018 dataset are allocated for training, while the remaining 20% are designated for testing. All approaches are trained using different parameters like epoch, batch size, and learning rate. The results indicate that combining ResNet-18 and MobileNet pre-trained networks using concatenation with an SVM classifier achieved the maximum accuracy of 92.87%.

In this paper, we have made a comparative analysis of ear recognition for constrained and unconstrained dataset. We have also analyzed the effect of image quality like illumination, occlusion, brightness, and pose variation in our work. We have addressed the problem of limited dataset for ear recognition task using deep learning-based approach. We aim at building a deep convolutional neural network model that does the task of classification for ear recognition task. Our model considers both constrained and unconstrained dataset of ear image for analysis purpose. In this paper, we have explored the benefits of deep learning and its different approached to enhance the performance ear based biometric systems where sample images are captured in constrained and unconstrained condition. This paper mainly focuses on creation of deep learning model and compared its performance on three different datasets (AWE, AMI, and IITD-II) of ear images by varying the network architecture.

Objective To develop a deep learning method for fast and accurate prediction of Specific Absorption Rate (SAR) distributions in the human head to support real-time hyperthermia treatment planning (HTP) of brain cancer patients. Approach We propose an encoder-decoder neural network with cross-attention blocks to predict SAR maps from brain electrical properties, tumor 3D isocenter coordinates and microwave antenna phase settings. A dataset of 201 simulations was generated using finite-element modeling by varying tissue properties, tumor positions, and antenna phases within a human head model equipped with a three-ring phased-array applicator. The model was trained and evaluated on this dataset using standard error metrics and structural similarity analysis. Main results On a held-out test set of 20 samples, the model achieved a mean root-mean-squared error (RMSE) of 3.3 W/kg and a mean absolute error (MAE) of 1.6 W/kg across the whole brain. In target regions, RMSE and MAE were 4.8 and 2.5 W/kg, respectively. The structural similarity index (SSIM) reached a mean value of 0.90, and the computation time was reduced from 10 min (simulation-based) to 4 s using our deep learning approach. The proposed method enables accurate, efficient SAR prediction for HTP in the brain, potentially supporting real-time HTP to optimize tumor temperature and improve clinical outcomes. Significance This work introduces a novel deep learning-based approach that significantly accelerates SAR calculation in HTP, enabling adaptive therapy strategies to improve treatment outcomes in hyperthermia.

Text classification is an essential component in a variety of applications of natural language processing. While the deep learning-based approach is becoming more popular, using vectors of word as an input for the models has proved to be a good way for the machine to learn the relation between words in a document. This paper proposes a solution for the text classification using hybrid deep learning approaches. Every existing deep learning approach has its own advantages and the hybrid deep learning model we are introducing is the combination of the superior features of CNN and LSTM models. The proposed models CNN-LSTM, LSTM-CNN show enhanced accuracy over another approach.

Assessment and feedback methods have important role in motivating students to take deep or surface approaches in their learning. Designing assessments with clear link to the teaching content and intended learning outcomes (ILOs), supports students’ approach toward deep learning. This study explored impact of the weekly short assessments with individual feedback in blended classes, on motivating students to take a deep learning approach in their studies. Constructive alignment framework was used to prepare the teaching content, in-class learning activities and assessment tasks. Series of short formative questions prepared with each question evolving around one teaching topic and its ILOs. Students’ responses to each question were followed by individual feedback and an opportunity to reflect on the received feedback to improve their work. Students’ perception towards the short formative assessment with individual feedback was collected using the Revised Study Process Questionnaire (R-SPQ-2F). Using quantitative methodology responses of the 90 Year 2 chemical and environmental engineering students were analysed. Results showed that there are significantly higher deep approach scores (Mean=2.92, SD=1.14) compared to surface approach (Mean=2.41, SD=1.15), with p-value=0.001. Students’ motives and strategies in taking either deep or surface approach is also explored. While factors of “self-satisfaction” and “interest in the course content” were the main motives for students to take deep learning approach, surface strategies such as “learning the examinable content” were remaining high. This research aims to contribute to development of an assessment method in engineering education to foster students’ deep learning, by developing their critical thinking and problem-solving skills.

BackgroundIn biomedical text mining, named entity recognition (NER) is an important task used to extract information from biomedical articles. Previously proposed methods for NER are dictionary- or rule-based methods and machine learning approaches. However, these traditional approaches are heavily reliant on large-scale dictionaries, target-specific rules, or well-constructed corpora. These methods to NER have been superseded by the deep learning-based approach that is independent of hand-crafted features. However, although such methods of NER employ additional conditional random fields (CRF) to capture important correlations between neighboring labels, they often do not incorporate all the contextual information from text into the deep learning layers.ResultsWe propose herein an NER system for biomedical entities by incorporating n-grams with bi-directional long short-term memory (BiLSTM) and CRF; this system is referred to as a contextual long short-term memory networks with CRF (CLSTM). We assess the CLSTM model on three corpora: the disease corpus of the National Center for Biotechnology Information (NCBI), the BioCreative II Gene Mention corpus (GM), and the BioCreative V Chemical Disease Relation corpus (CDR). Our framework was compared with several deep learning approaches, such as BiLSTM, BiLSTM with CRF, GRAM-CNN, and BERT. On the NCBI corpus, our model recorded an F-score of 85.68% for the NER of diseases, showing an improvement of 1.50% over previous methods. Moreover, although BERT used transfer learning by incorporating more than 2.5 billion words, our system showed similar performance with BERT with an F-scores of 81.44% for gene NER on the GM corpus and a outperformed F-score of 86.44% for the NER of chemicals and diseases on the CDR corpus. We conclude that our method significantly improves performance on biomedical NER tasks.ConclusionThe proposed approach is robust in recognizing biological entities in text.

Remote sensing is a key tool to derive glacier surface velocities but existing mapping methods, such as cross-correlation techniques, can fail where surface properties change temporally or where large velocity variations occur spatially. High-resolution datasets, such as UAV imagery, offer a promising solution to tackle these issues and to study small-scale glacier dynamics, but new workflows are required to handle such data. Therefore, we tested the potential of new deep learning-based image-matching algorithms for deriving glacier surface velocities across the ablation area of a glacier with strong spatial variability in surface velocities (<5 m/yr to >100 m/yr) and substantial changes in surface properties between image acquisitions. For a thorough comparison of state-of-the-art methods and sensors, we applied three different techniques (cross-correlation using geoCosiCorr3D, feature tracking with ORB using SeaIceDrift and the new deep learning-based method using ICEpy4D) and three different platforms (Sentinel-2, PlanetScope, UAVs) to estimate glacier surface velocities. Results showed lowest errors for velocities derived with the deep learning-based approach applied to UAV imagery (RMSE = 2.17 m/yr, R2 = 0.99), followed by cross-correlation using Sentinel-2 images (RMSE = 21.0 m/yr, R2 = 0.59) and the deep learning-based approach with PlanetScope data (RMSE = 21.28 m/yr, R2 = 0.36). Cross-correlation with geoCosiCorr3D resulted in comparably high errors with the UAV dataset (RMSE = 36.22 m/yr, R2 = 0.24), whereas ORB-based feature tacking showed lowest performance with all sensors. Spatial patterns of computed velocities indicate that applying existing cross-correlation methods for areas with regular displacements or low glacier velocities yields suitable results on UAV data, but innovative deep learning-based approaches are required for resolving rapid changes in velocities or in surface properties. This novel method benefits from improved keypoint detection and matching through training using neural networks and data characterized by challenging geometries, outlier minimization and more robust descriptors by applying cross-attention layers. We conclude that continued development of deep learning-based feature tracking approaches for glacier velocity computations may substantially improve UAV-based velocity derivations applied to challenging situations. This method is able to deliver reliable displacement data in situations where traditional methods fail, which implies a new level of detail in understanding and interpreting glacier dynamics.

This paper focuses on the real-time remaining useful life (RUL) estimation of SAC305, SAC105 and SnPb solder alloy PCBs subjected to combined temperature and vibration loads. The RUL estimation of the packages on the PCB were carried out using a combined physics of failure and deep learning approaches for different operating conditions. The test boards used in this study are of same configuration for all the three solder materials and it consists of a multilayer FR4 configuration with JEDEC standard dimensions. The failure predictions and feature vector identifications are carried out using the strain gauge signals attached at the back of the PCB. The strain signals are analyzed both in the time and frequency domain to identify the different feature vectors that can predict failure of the packages as the number of drop increases. Principal component analysis is used as the pattern recognition and data reduction technique for the time and frequency domain data of the strain signals. Frequency components including and excluding the natural frequency of the test boards were used to identify the different patterns of before and after failure strain signals. The remaining useful life estimations are very useful in improving the efficiency and proactively helps to schedule maintenance effectively. The use of deep learning helps to models complex systems with multiple parameters involving nonlinear behaviors. In this paper the feature vectors identified from different operating conditions are modelled using a combined physics of failure and deep learning-based approach to estimate the remaining useful life of the packages. The changes in the material characteristics of the solders with different operating conditions are also modelled to the Long Short-term Memory (LSTM) deep learning model with the feature vectors to predict the failure of the packages. A regression model to predict the failure is also modelled to predict the failure based on the loading and material characteristics of the solders. LSTM models for each solder materials for multiple use-cases are modeled, and combined models involving different acceleration levels are also modeled.

Knowing which residues of a protein are important for its function is of paramount importance for understanding the molecular basis of this function and devising ways of modifying it for medical or biotechnological applications. Due to the difficulty in detecting these residues experimentally, prediction methods are essential to cope with the sequence deluge that is filling databases with uncharacterized protein sequences. Deep learning approaches are especially well suited for this task due to the large amounts of protein sequences for training them, the trivial codification of this sequence data to feed into these systems, and the intrinsic sequential nature of the data that makes them suitable for language models. As a consequence, deep learning-based approaches are being applied to the prediction of different types of functional sites and regions in proteins. This review aims to give an overview of the current landscape of methodologies so that interested users can have an idea of which kind of approaches are available for their proteins of interest. We also try to give an idea of how these systems work, as well as explain their limitations and high dependence on the training set so that users are aware of the quality of expected results.