Smirnova, Nadezhda S

Demertzis K., Iliadis L., Pimenidis E., Tziritas N., Koziri M., Kikiras P.

Maximizing the production process in modern industry, as proposed by Industry 4.0, requires extensive use of Cyber-Physical Systems (CbPS). Artificial intelligence technologies, through CbPS, allow monitoring of natural processes, making autonomous, decentralized and optimal decisions. Collection of information that optimizes the effectiveness of decisions, implies the need for big data management and analysis. This data is usually coming from heterogeneous sources and it might be non-interoperable. Big data management is further complicated by the need to protect information, to ensure business confidentiality and privacy, according to the recent General Data Protection Regulation - GDPR. This paper introduces an innovative holistic Blockchained Adaptive Federated Auto Meta Learning Big Data and DevOps Cyber Security Architecture in Industry 4.0. The aim is to fill the gap found in the ways of handling and securing industrial data. This architecture, combines the most modern software development technologies under an optimal and efficient framework. It successfully achieves the prediction and assessment of threat-related conditions in an industrial ecosystem, while ensuring privacy and secrecy.

Psathas A.P., Iliadis L., Papaleonidas A., Bountas D.

The daily growth of computer networks usage increases the need to protect users from malware and other threats. This paper, presents a hybrid Intrusion Detecting System (IDS) comprising of a 2-Dimensional Convolutional Neural Network (2-D CNN), a Recurrent Neural Network (RNN) and a Multi-Layer Perceptron (MLP) for the detection of 9 Cyber Attacks versus normal flow. The timely Kitsune Network attack dataset was used in this research. The proposed model achieved an overall accuracy of 92.66%, 90.64% and 90.56% in the train, validation and testing phases respectively. The typical five classification indices Sensitivity, Specificity, Accuracy, F1-Score and Precision were calculated following the “One-Versus-All Strategy”. Their values clearly support the fact that the model can generalize and that it can be used as a prototype for further research on network security enhancement.

Herzog N.J., Magoulas G.D.

Advances in neural networks and deep learning have opened a new era in medical imaging technology, health care data analysis and clinical diagnosis. This paper focuses on the classification of MRI for diagnosis of early and progressive dementia using transfer learning architectures that employ Convolutional Neural Networks-CNNs, as a base model, and fully connected layers of Softmax functions or Support Vector Machines-SVMs. The diagnostic process is based on the analysis of the neurodegenerative changes in the brain using segmented images of brain asymmetry, which has been identified as a predictive imaging source of early dementia. Results from 300 independent simulation runs on a set of four binary and one multiclass MRI classification tasks illustrate that transfer learning of CNN-based models equipped with SVM output layer is capable to produce better performing models within a few training epochs compared to commonly used transfer learning architectures that combine CNN pretrained models with fully connected Softmax layers. However, experimental findings also confirm that longer training sessions appear to compensate for the shortcomings of the fully connected Softmax layers in the long term. Diagnosis of early dementia on unseen patients’ brain asymmetry MRI data reached an average accuracy of 90.25% with both transfer learning architectures, while progressive dementia was promptly diagnosed with an accuracy that reached 95.90% using a transfer learning architecture that has the SVM layer.

Pirie C., Moreno-Garcia C.F.

Inspection engineering is a highly important field in the Oil & Gas sector for analysing the health of offshore assets. Corrosion, a naturally occurring phenomenon, arises as a result of a chemical reaction between a metal and its environment, causing it to degrade over time. Costing the global economy an estimated US $2.5 Trillion per annum, the destructive nature of corrosion is evident. Following the downturn endured by the industry in recent times, the need to combat corrosion is escalated, as companies look to cut costs by increasing efficiency of operations without compromising critical processes. This paper presents a step towards automating solutions for real-time inspection using state-of-the-art computer vision and deep learning techniques. Experiments concluded that there is potential for the application of computer vision in the inspection domain. In particular, Mask R-CNN applied on the original images (i.e. without any form of pre-processing) was found to be most viable solution, with the results showing a mAP of 77.1%.

Abdullakutty F., Elyan E., Johnston P.

Face recognition has recently become widespread in security applications. Although advancing technology has improved the performance of these systems, they are still prone to various attacks, including spoofing. The inherent feature extraction capability of machine learning techniques and deep neural networks has facilitated more accurate performance in spoofing detection. However, challenges still remain in the generalisation of these methods. One significant challenge is training dataset limitation in terms of size and variance. This paper investigates how different train/test ratios and variance in training data affect model performance with the NUAA dataset for spoofing detection. We show how using different splits of this dataset results in different models with different performances. We also open up new research directions by demonstrating how the problem of generalisation can be neatly demonstrated with an existing manageable dataset.

Wu B., Waschneck B., Mayr C.

In this work, we propose a method based on attention to quantize convolutional neural networks (CNNs) to run on low-precision (binary and multi-bit). Intuitively, high-quality pictures are very conducive to distinguishing objects. However, even in low-quality black-and-white photos (analogous to low-precision), various features can also be well distinguished and the content is easily understood. Based on this intuition, we introduce an attention-based block called squeeze-and-threshold (ST) to adjust different features to different ranges and learn the best threshold to distinguish (quantize) them. Furthermore, to eliminate the extra calculations caused by the ST block in the inference process, we propose a momentum-based method to learn the inference threshold during the training stage. Additionally, with the help of ST block, our quantization approach is faster and takes less than half the training epochs of prior multi-bit quantization works. The experimental results on different datasets and networks show the versatility of our method and demonstrate state-of-the-art performance.

Rutigliano D., Boracchi G., Invernizzi P., Sozio E., Alippi C., Binetti S.

The Optical Time Domain Reflectometer (OTDR) is an optoelectronic instrument used to characterize an optical fiber using the measure of scattered or reflected light from points along the fiber. The resulting signal, namely the OTDR trace, is commonly used to identify and localize possible critical events in the fiber. In this work we address the problem of automatically detecting optical events in OTDR traces, and present the first 1D object-detection neural network for optical trace analysis. Our approach takes inspiration from a successful object detection network in images, the Faster R-CNN, which we adapt to time series domain. The proposed network can both classify and localize many optical events along an input trace. Our results show that the proposed solution is more accurate than existing software currently analyzing OTDR traces, improving the mean average precision score by $$27.43\%$$ . In contrast with existing solutions that are not able to distinguish many types of events, our algorithm can be trained in an end-to-end manner to detect potentially any type of optic event. Moreover, our network has been deployed on embedded OTDR devices to be executed in real-time.

Giannou O., Giannou A.D., Zazara D.E., Kleinschmidt D., Mummert T., Stüben B.O., Kaul M.G., Adam G., Huber S., Pavlidis G.

Machine learning techniques have provided a technological evolution in medicine and especially in the field of medical imaging. The aim of this study is to firstly compare multiple transfer learning architecture models such as MobilleNetVn (n = 1, 2), InceptionVn (n = 1, 2, 3, 4), InceptionResNet, VGG16 and NasNetMobile and provide a final performance estimation, using a variety of metrics, secondly, propose an efficient and accurate classifier for liver cancer trait detection and prediction, and thirdly, develop a mobile application which uses the proposed model to classify liver cancer traits into various categories in real-time. Magnetic Resonance Images (MRI) of mouse liver cancer of different origin were used as input datasets for our experiments. However, the required memory by the deployed Convolutional Neural Network (CNN) models on smart mobile devices or embedded systems for real time applications, is an issue to be addressed. Here, all the baseline pre-trained CNN models on the ImageNet dataset were trained on a dataset of MRI images of mice of different genetic background with genetically- or chemically- induced hepatocellular cancer. We present and compare all main metric values for each model such as accuracy, cross entropy, f-score, confusion matrix for various types of classification. Data analysis verifies that the proposed optimized architecture model for this task of liver cancer trait classification and prediction, the MV1-LCCP, shows a suitable performance in terms of memory utilization and accuracy, suitable to be deployed in the mobile Android application, which is also developed and presented in this paper.

Nandi A., Xhafa F., Subirats L., Fort S.

Emotions of learners are crucial and important in e-learning as they promote learning. To investigate the effects of emotions on improving and optimizing the outcomes of e-learning, machine learning models have been proposed in the literature. However, proposed models so far are suitable for offline mode, where data for emotion classification is stored and can be accessed boundlessly. In contrast, when data arrives in a stream, the model can see the data once and real-time response is required for real-time emotion classification. Additionally, researchers have identified that single data modality is incapable of capturing the complete insight of the learning experience and emotions. So, multi-modal data streams such as electroencephalogram (EEG), Respiratory Belt (RB), electrodermal activity data (EDA), etc., are utilized to improve the accuracy and provide deeper insights in learners’ emotion and learning experience. In this paper, we propose a Real-time Multimodal Emotion Classification System (ReMECS) based on Feed-Forward Neural Network, trained in an online fashion using the Incremental Stochastic Gradient Descent algorithm. To validate the performance of ReMECS, we have used the popular multimodal benchmark emotion classification dataset called DEAP. The results (accuracy and F1-score) show that the ReMECS can adequately classify emotions in real-time from the multimodal data stream in comparison to the state-of-the-art approaches.

Abid M., Lefebvre G.

Geomagnetic field fingerprinting is an attractive alternative to WiFi and Bluetooth fingerprinting since the magnetic field is omnipresent and independent of any infrastructure. Recently, Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have been extensively used to provide fingerprinting solutions for indoor positioning based on magnetic data. Yet, no fairly comparative study has been conducted to determine which type of networks is more likely to recognize magnetic sequence patterns. In this study, we propose a CNN using Recurrence Plots (RPs) as sequence fingerprints as well as advanced RNNs treating magnetic sequences as fingerprints. We used the same real-world data in an indoor environment to build and fairly evaluate the proposed systems. Our findings show that RNNs clearly outperform the RP-based CNN, yet results in higher prediction latencies. Overall, promising positioning performances and smooth trajectory estimates are achieved for pedestrian path tracking due to approaching the indoor localization problem from a regression perspective.

Shahra E.Q., Wu W., Basurra S., Rizou S.

Maintaining high water quality is the main goal for water management planning and iterative evaluation of operating policies. For effective water monitoring, it is crucial to test a vast number of drinking water samples that is time-consuming and labour-intensive. The primary objective of this study is to determine, with high accuracy, the quality of drinking water samples by machine learning classification models while keeping computation time low. This paper aims to investigate and evaluate the performance of two supervised classification algorithms, including artificial neural network (ANN) and support vector machine (SVM) for multiclass water classification. The evaluation uses the confusion matrix that includes all metrics ratios, such as true positive, true negative, false positive, and false negative. Moreover, the overall accuracy and f1-score of the models are evaluated. The results demonstrate that ANN outperformed the SVM with an overall accuracy of 94% in comparison to SVM, which shows an overall accuracy of 89%.

Symeonidis C., Nousi P., Tosidis P., Tsampazis K., Passalis N., Tefas A., Nikolaidis N.

The performance of supervised deep learning algorithms depends significantly on the scale, quality and diversity of the data used for their training. Collecting and manually annotating large amount of data can be both time-consuming and costly tasks to perform. In the case of tasks related to visual human-centric perception, the collection and distribution of such data may also face restrictions due to legislation regarding privacy. In addition, the design and testing of complex systems, e.g., robots, which often employ deep learning-based perception models, may face severe difficulties as even state-of-the-art methods trained on real and large-scale datasets cannot always perform adequately as they have not adapted to the visual differences between the virtual and the real world data. As an attempt to tackle and mitigate the effect of these issues, we present a method that automatically generates realistic synthetic data with annotations for a) person detection, b) face recognition, and c) human pose estimation. The proposed method takes as input real background images and populates them with human figures in various poses. Instead of using hand-made 3D human models, we propose the use of models generated through deep learning methods, further reducing the dataset creation costs, while maintaining a high level of realism. In addition, we provide open-source and easy to use tools that implement the proposed pipeline, allowing for generating highly-realistic synthetic datasets for a variety of tasks. A benchmarking and evaluation in the corresponding tasks shows that synthetic data can be effectively used as a supplement to real data.

Stylianou-Nikolaidou S., Vernikos I., Mathe E., Spyrou E., Mylonas P.

The problem of human activity recognition (HAR) has been increasingly attracting the efforts of the research community, having several applications. In this paper we propose a multi-modal approach addressing the task of video-based HAR. Our approach uses three modalities, i.e., raw RGB video data, depth sequences and 3D skeletal motion data. The latter are transformed into a 2D image representation into the spectral domain. In order to extract spatio-temporal features from the available data, we propose a novel hybrid deep neural network architecture that combines a Convolutional Neural Network (CNN) and a Long-Short Term Memory (LSTM) network. We focus on the tasks of recognition of activities of daily living (ADLs) and medical conditions and we evaluate our approach using two challenging datasets.

Ieracitano C., Mammone N., Paviglianiti A., Morabito F.C.

Electrospinning is a manufacturing technique used to produce nanofibers for engineering applications. This process depends on several control parameters (such as solution concentration, applied voltage, flow rate, tip-to-collector distance) whose variations during the experiments may lead to the formation of defective nanofibers (D-NF) along with non-defective nanofibers (ND-NF). D-NF present either with impurities or morphological defects that prevent their practical use in nanotechnology. In this context, here, a data augmentation based machine learning approach is proposed to classify Scanning Electron Microscope (SEM) images in two classes (i.e., D-NF vs. ND-NF). To this end, a custom Convolutional Neural Network (CNN) is developed to perform the binary classification task, achieving an accuracy rate up to 93.85%. Moreover, the explainability of the proposed CNN is also explored by means of an occlusion sensitivity analysis in order to investigate which area of the SEM image mostly contributes to the classification task. The achieved encouraging findings stimulate the use of the proposed framework in industrial applications.

Vrochidis A., Dimitriou N., Krinidis S., Panagiotidis S., Parcharidis S., Tzovaras D.

Over the last decade, the volume of videos available on the web has increased exponentially. In order to help users cope with the ever-growing video volume, recommendation systems have emerged that can provide personalized suggestions to users based on their past preferences and relevant online metrics. However, such approaches require user profiling, which raises privacy issues while often providing delayed suggestions as various metrics have to be firstly collected such as ratings and number of views. In this paper, we propose a system specifically targeting video content generated in a conference event, where a series of talks and presentations are held and a separate video for each is recorded. Through audience analysis, our system is able to predict the online views of each video and thus recommend the most popular videos to users. This way, online users don’t have to search through all the videos of a conference event thus saving time while not missing the most impactful videos. The proposed system employs several complementary techniques for audience analysis based on video and audio streams. Experimental evaluation of real data demonstrates the potential of the proposed approach.