Πλοήγηση ανά Επιβλέπων "Koutsopoulos, Iordanis"

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω

Τώρα δείχνει 1 - 10 από 10

Attack detection and digital forensics analysis in mobile Ad Hoc Networks (MANETs)
(15-02-2023) Δρούγκα, Σοφία; Drougka, Sofia; Athnes University of Economics and Business, Department of Informatics; Polyzos, George; Gkritzalis, Dimitrios; Koutsopoulos, Iordanis
Σκοπός της παρούσας πτυχιακής εργασίας είναι η εξέταση διαφόρων τύπων δικτυακών επιθέσεων και πιο συγκεκριμένα εκείνων που στοχεύουν σε Κινητά Ad Hoc Δίκτυα (MANETs), καθώς και η έγκαιρη και αποτελεσματική διερεύνηση και ανίχνευση των επιθέσεων αυτών. Τα MANETs είναι κατανεμημένα ασύρματα κινητά Ad Hoc δίκτυα που χρησιμοποιούνται σε πολλές σύγχρονες υπηρεσίες, συμπεριλαμβανομένης της τηλεϊατρικής, της άμυνας, της πλοήγησης οχημάτων και του Διαδικτύου των Πραγμάτων χάρη στα οφέλη που προσφέρουν σε σύγκριση με τα συμβατικά δίκτυα. Επιπλέον, η έρευνα έχει ως κύριο στόχο της την ανίχνευση, καθώς και τον εντοπισμό συγκεκριμένων τύπων επιθέσεων δικτύου με έγκαιρο και αποτελεσματικό τρόπο. Η ανίχνευση επιθέσεων πραγματοποιείται μετά από εκτεταμένη μελέτη της συμπεριφοράς των κόμβων που υπάρχουν στο δίκτυο, τόσο όταν βρίσκεται σε κανονική κατάσταση λειτουργίας όσο και όταν ένας ή περισσότεροι κόμβοι πραγματοποιούν επίθεση. Για όλους τους παραπάνω λόγους, στην παρούσα εργασία διεξάγεται μια ολοκληρωμένη μελέτη που αφορά το σχεδιασμό διαφόρων σεναρίων επίθεσης, αλλά και την υλοποίηση διαφόρων πειραμάτων στο δίκτυο. Πιο συγκεκριμένα, τα πειράματα αυτά αφορούν δύο κατηγορίες δικτυακών επιθέσεων, οι οποίες λαμβάνουν χώρα στο Επίπεδο Μεταφοράς του μοντέλου OSI, χρησιμοποιώντας το πρωτόκολλο TCP και UDP.Η εισαγωγή και τα επόμενα κεφάλαια αυτής της διατριβής περιλαμβάνουν περισσότερες λεπτομέρειες σχετικά με τις επιθέσεις άρνησης υπηρεσίας και τις κατανεμημένες επιθέσεις άρνησης υπηρεσίας. Επίσης, θα αξιολογηθεί η σημασία του πρωτοκόλλου OLSR, των MANETs και της λειτουργίας τους. Αξίζει να σημειωθεί ότι τα πειράματα πραγματοποιήθηκαν χρησιμοποιώντας το εργαλείο CORE Emulator, το οποίο είναι ένα λογισμικό για την δημιουργία εικονικών δικτύων και λειτουργεί ως εξομοιωτής πραγματικών δικτύων. Επιπλέον, η ανάλυση και ταυτοποίηση των επιθέσεων πραγματοποιήθηκε με τη βοήθεια μεθοδολογιών ψηφιακής εγκληματολογίας, καθώς και με τη βοήθεια του περιβάλλοντος ELK Stack. Στο τελευταίο κεφάλαιο της εργασίας, παρουσιάζεται μια πρόταση για περαιτέρω εμβάθυνση στην έρευνα όσον αφορά την δημιουργία ενός συστήματος ανίχνευσης δικτυακών επιθέσεων βασισμένο στη μηχανική μάθηση.
Computer vision on piping and instrumentation diagrams: towards the identification of their components
(08-12-2021) Androvitsanea, Anna; Ανδροβιτσανέα, Άννα; Athens University of Economics and Business, Department of Informatics; Vassalos, Vasilios; Koutsopoulos, Iordanis
Piping and Instrumentation Diagrams (P&IDs) are schematic representations of equipment, pipelines, instrumentation, and control systems. They appear in process environments, such as Oil Refineries, Chemical Plants, Paper Mills, and Cement Plants, etc. The identification of each element constituting a P&ID, along with the way they are interconnected, is an important task that has not been automated yet. In this work we study a methodology and develop the respective algorithm towards the identification of these components. This identification aims to the classification of the elements based on their representation as images as well as to the identification and translation of the codes included in the diagrams.In order to achieve this goal a combination of methods are employed. Using the OpenCV library the outlines of the P&ID are calculated. An algorithm is developed, which based on the coordinates of the outlines, delivers snapshots of the elements constituting the P&ID. In the sequel, these elements are classified by a suitably designed classifier, to one out of 53 classes. The classifier is a convolutional neural network (CNN), implemented using the TensorFlow and Keras libraries, which was trained on a data set of 2970 images that belong to one out of 53 classes.Textual information contained in the P&ID are identified, using the pytesseract library and stored into an array. Then, they are passed to an algorithm that implements the ANSI/ISA-5.1.-1984 (R1992) standards and deciphers the textual tags, by providing as output the name, function, modifier etc. of each element. The model is able to successfully identify an image and attribute it to the right class, which is a great step towards solving the challenging problem of the identification of the elements constituting a P&ID.cPiping and Instrumentation Diagrams (P&IDs) are schematic representations of equipment, pipelines, instrumentation, and control systems. They appear in process environments, such as Oil Refineries, Chemical Plants, Paper Mills, and Cement Plants, etc. The identification of each element constituting a P&ID, along with the way they are interconnected, is an important task that has not been automated yet. In this work we study a methodology and develop the respective algorithm towards the identification of these components. This identification aims to the classification of the elements based on their representation as images as well as to the identification and translation of the codes included in the diagrams.In order to achieve this goal a combination of methods are employed. Using the OpenCV library the outlines of the P&ID are calculated. An algorithm is developed, which based on the coordinates of the outlines, delivers snapshots of the elements constituting the P&ID. In the sequel, these elements are classified by a suitably designed classifier, to one out of 53 classes. The classifier is a convolutional neural network (CNN), implemented using the TensorFlow and Keras libraries, which was trained on a data set of 2970 images that belong to one out of 53 classes.Textual information contained in the P&ID are identified, using the pytesseract library and stored into an array. Then, they are passed to an algorithm that implements the ANSI/ISA-5.1.-1984 (R1992) standards and deciphers the textual tags, by providing as output the name, function, modifier etc. of each element. The model is able to successfully identify an image and attribute it to the right class, which is a great step towards solving the challenging problem of the identification of the elements constituting a P&ID.
Continual learning in encoder-decoder computer vision architectures
(29-11-2024) Βεχλίδης, Κωνσταντίνος; Vechlidis, Konstantinos; Athens University of Economics and Business, Department of Informatics; Toumpis, Stavros; Pavlopoulos, Ioannis; Koutsopoulos, Iordanis
Η συνεχής μάθηση (Continual Learning) επιτρέπει στα μοντέλα να προσαρμόζονται σε νέα δεδομένα, διατηρώντας παράλληλα τη γνώση που είχαν μάθει προηγουμένως, αντιμετωπίζοντας έτσι το πρόβλημα της καταστροφικής λήθης (catastrophic forgetting). Η παρούσα διατριβή διερευνά την εφαρμογή της συνεχούς μάθησης σε αρχιτεκτονικές κωδικοποιητή-αποκωδικοποιητή, εστιάζοντας σε προβλήματα υπολογιστικής όρασης, όπως η δημιουργία λεζάντας εικόνας (image captioning). Οι αρχιτεκτονικές κωδικοποιητή-αποκωδικοποιητή επιλέγονται κυρίως σε προβλήματα που απαιτούν τη μετατροπή δομημένων δεδομένων εισόδου σε άλλη δομή ή μέσο, όπως η παραγωγή κειμένου από μια εικόνα. Τέτοια προβλήματα συχνά περιλαμβάνουν εξελισσόμενες κατανομές δεδομένων, καθιστώντας τη συνεχή μάθηση σε αυτές τις αρχιτεκτονικές απαραίτητη για τη διατήρηση της απόδοσης και της προσαρμοστικότητας. Υιοθετούμε το σενάριο της μάθησης μέσω σταδιακής επαύξησης κλάσεων (class-incremental learning), όπου νέες κλάσεις εισάγονται σταδιακά και το μοντέλο πρέπει να μάθει να ταξινομεί τις νέες κλάσεις χωρίς να ξεχνά τις προηγούμενες. Για παράδειγμα, σε ένα πρόβλημα ταξινόμησης εικόνων, ένα μοντέλο που έχει αρχικά εκπαιδευτεί να ταξινομεί κατηγορίες όπως «γάτα» και «σκύλος» μπορεί αργότερα να χρειαστεί να μάθει σταδιακά να ταξινομεί νέες κατηγορίες, όπως «πουλί» ή «ψάρι», χωρίς να χάσει την ικανότητά του να αναγνωρίζει σωστά τη «γάτα» και τον «σκύλο» σε μελλοντικές εισόδους. Για να αντιμετωπίσουμε αυτό το σενάριο, χρησιμοποιούμε τον αλγόριθμο Gradient Episodic Memory (GEM), μια διάσημη τεχνική που βασίζεται στην επανάληψη και τη βελτιστοποίηση. Ο GEM μετριάζει την καταστροφική λήθη, αποθηκεύοντας δεδομένα προηγούμενων εργασιών στη μνήμη και περιορίζοντας τις κλίσεις των παραγώγων κατά τη διάρκεια νέων φάσεων μάθησης, ώστε να αποφεύγονται παρεμβολές με τις γνώσεις που είχαν αποκτηθεί προηγουμένως. Σε αυτή την εργασία, επιλέγουμε τον GEM λόγω της ικανότητάς του να χειρίζεται τόσο την επανάληψη όσο και τη βελτιστοποίηση σε ένα ενοποιημένο πλαίσιο. Ο GEM εξασφαλίζει ότι οι εργασίες που έχουν διδαχθεί προηγουμένως δεν ξεχνιούνται καθώς το μοντέλο εκτίθεται σε νέες εργασίες, γεγονός ιδιαίτερα κρίσιμο στη μάθηση μέσω σταδιακής επαύξησης κλάσεων για τη δημιουργία λεζάντας εικόνας, όπου κάθε νέα εργασία εισάγει νέα λεξιλογικά σημεία (κλάσεις). Επιπλέον, προτείνουμε μια νέα έκδοση του GEM, που χρησιμοποιεί μάσκα, για να αντιμετωπίσουμε τη συνεχή αύξηση των παραμέτρων του μοντέλου λόγω των σταδιακών ενημερώσεων του λεξιλογίου. Εξ' όσων γνωρίζουμε, αυτή είναι η πρώτη προσπάθεια εφαρμογής του GEM στο πλαίσιο της δημιουργίας λεζάντας εικόνας. Μέσω πειραμάτων, αποδεικνύουμε ότι ο GEM υπερτερεί έναντι άλλων σύγχρονων τεχνικών συνεχούς μάθησης που έχουν εφαρμοστεί στη δημιουργία λεζάντας εικόνας, όπως το Feature Distillation (FD) και το Learning without Forgetting (LwF).
Identifying problematic gambling behaviour
(2021) Markakis, Christos; Μαρκάκης, Χρήστος; Athens University of Economics and Business, Department of Informatics; Vrontos, Ioannis; Vassalos, Vasilios; Koutsopoulos, Iordanis
In recent years more and more people are getting exposed to internet gambling which makes it necessary to develop methods of monitoring their behaviour. This thesis research methods that can identify problematic gambling behaviour in an unsupervised manner. We utilize real behavioural tracking data from online gambling activity to group players to problem gambling groups with similar gambling behaviour using clustering. Furthermore, we propose an approach for categorizing the players into groups of problem gambling according to their severity and for training machine learning models to be able to classify them. This research is implemented on data concerning players that have significant account activity. The results of clustering show that among the clustering techniques that were implemented, K-means is the most capable one to identify distinct groups, with different characteristics and some of them indicating problematic gambling behaviour. Additionally, the second approach distinguishes players into problem gambling groups of increasing severity and between the classification techniques that were tested, Neural Network models with oversampling were the best-performing ones in classifying players into problem gambling groups. The results of this thesis can be used to group players with similar problematic gambling behaviour and identify them in a different set of players in order to protect the players and the companies from the destructive consequences of gambling.
Investigating hallucinations in AI-based text generation using semantic entropy
(07-02-2025) Αγγελίδης, Αναστάσιος; Angelidis, Anastasios; Athens University of Economics and Business, Department of Informatics; Toumpis, Stavros; Pavlopoulos, Ioannis; Koutsopoulos, Iordanis
Τα Μεγάλα Γλωσσικά Μοντέλα (LLMs) έχουν φέρει επανάσταση στη Γενετική Φυσικής Γλώσσας (NLG), αλλά πάσχουν από παραισθήσεις—παραγόμενο περιεχόμενο που είναι πραγματολογικά λανθασμένο ή μη πιστό στο αρχικό υλικό. Αυτή η διπλωματική εργασία διερευνά τις επινοήσεις, μια συγκεκριμένη υποκατηγορία παραισθήσεων, χρησιμοποιώντας ένα πλαίσιο σημασιολογικής εντροπίας. Μέσω ομαδοποίησης σημασιολογικά ισοδύναμων απαντήσεων και υπολογισμού της εντροπίας, το πλαίσιο εντοπίζει αβέβαιες εξόδους που είναι πιθανό να αποτελούν επινοήσεις. Μια βασική τροποποίηση αντικαθιστά την υπολογιστικά απαιτητική αξιολόγηση σημασιολογικά ισοδύναμων απαντήσεων με LLMs με ελαφριά μοντέλα Transformer, προσαρμοσμένα για Συμπερασματική Φυσικής Γλώσσας (NLI). Τα πειραματικά αποτελέσματα σε σύνολα δεδομένων αξιολόγησης (TriviaQA, SQuAD, SVAMP και NQ Open) δείχνουν βελτιωμένη ακρίβεια, κλιμάκωση και αποδοτικότητα.
New algorithms and practical implementation issues in federated learning
(10/26/2021) Tsouparopoulos, Thomas; Τσουπαρόπουλος, Θωμάς; Athens University of Economics and Business, Department of Informatics; Toumpis, Stavros; Polyzos, George C.; Koutsopoulos, Iordanis
Federated Learning (FL) is a distributed Machine Learning paradigm that aims to traina central neural network model, on decentralized data located on edge devices. This isaccomplished by training local instances of the global model on the edge devices andthen aggregating their local solutions on a server; without accessing their private localdata. In this thesis, we make contributions in two fronts. First, in the practical implementationfront, we implement the most widely used FL algorithm, Federated Averaging (FedAvg),over a wireless network, using Raspberry Pi devices as clients, with the purpose of capturingthe cost of the negative externalities on the training quality. Second, in terms of advancing knowledge in FL algorithms, we propose a novel training pipeline for Federated learning that is based on the idea of Transfer Learning and uses the generative properties of Generative Adversarial Networks (GANs) to feed information about the decentralized datasets, to all the clients’ local networks. Finally, we provide some early encouraging results that indicate an improvement of 3,9 in central model accuracy on the test dataset, without increasing the communication overhead of the Federate Learning training procedure.
A recommender system for smart energy grids
(30-04-2021) Athanasakou, Alexandra; Αθανασάκου, Αλεξάνδρα; Athens University of Economics and Business, Department of Informatics; Polyzos, George; Siris, Vasilios; Koutsopoulos, Iordanis
In recent years, the everyday routine causes the lack of energy resources because of the ignorance of the impact of our actions. But what if we start to handle energy efficiently? Recommender Systems are here to help us. As Recommender Systems developed for energy efficiency are among the most popular Smart City goals, there are many different implementations in the literature. Among others, Matrix Factorization has been consolidated as the best performing approach especially due to the huge amounts of sensor data generated at city scale in many applications. This thesis aims to develop a Recommender System for Smart Energy Grids that offers energy efficiency tips by handling energy consumption data and collecting explicit feedback from households. Our approach consists of a set of Matrix Factorization models using SVD algorithm which is trained with the help of Clustering, and especially K-Means algorithm, in order to offer predictions fast and memory-efficiently compared to a single Matrix Factorization model. Moreover, in order to achieve higher prediction accuracy, the missing ratings of households are filled with the mean rating of the cluster they belong to.At the same time, the final recommendation is provided from a set of “Real-Time rules” that aim to filter the predictions list from Matrix Factorization models concerning the recommendation time frame, the devices used by the target household, and its interests. Right after, the Recommender System observes the real-time consumption of the target household’s electrical devices appeared in the filtered prediction list to find possible increased consumption based on the household’s past behavior. In summary, this thesis highlights the importance of Clustering in Recommender Systems along with Matrix Factorization in order to offer accurate predictions in a fast and memory-efficiently way and introduces a set of “Real-Time rules” for providing the final, personalized recommendation.
Resource allocation, content recommendations and online learning mechanisms for mobile edge computing
(05/23/2022) Χατζηελευθερίου, Λίβια-Έλενα; Athens University of Economics and Business, Department of Informatics; Polyzos, George; Toumpis, Stavros; Iosifidis, George; Dimakis, Antonios; Siris, Vasileios; Stamoulis, Georgios; Koutsopoulos, Iordanis
The Mobile Edge Computing (MEC) paradigm brings computing and cache capacity resources in the proximity of users. It gives rise to a new ecosystem of services, such as Augmented Reality (AR) ones, while reducing the latency that is experienced by users and lowering network service costs. The main challenges that MECfaces are related to the scarcity of resources at the network edge, the unpredictability of important system parameters, such as traffic, content and computation demand, and the ultra-low latency requirements that must be satisfied.In this thesis we deal with the challenges above, towards the optimization of two MEC goals: content delivery and real-time analytics at the edge of the network. We present resource allocation mechanisms and methods that automate the resource allocation, for fifth-generation (5G), Beyond-5G (B5G) and sixth-generation (6G) communication systems, accounting for edge resources such as caches, computational resources of mobiledevices and edge servers, bandwidth and energy. We tackle both offline and Online Learning (OL) instances of optimization problems that span content recommendationsand caching, user association and allocation of computing resources. We use a variety of mathematical tools to solve these problems, such as combinatorial optimization, convex optimization and Online Convex Optimization (OCO), a special case of OL. We analyse and we exploit the structural properties of the formulated optimization problems, either by designing algorithms ex novo, or by adapting existing techniques to our settings. We provide cost-efficient, fast and elegant solutions with provable performance guarantees, for a variety of important problems that arise within the MEC context. Overall, this Ph.D. thesis tackles a set of important optimization problems that arise in the context of edge computing and networking. We present novel problem formulations and algorithms that lead to solutions with provable performance guarantees, bringing the Mobile Edge Computing (MEC) paradigm a step closer to its practical realization.
Social networks and their impact on wireless networks
Noutsi, Evgenia; Νούτση, Ευγενία; Athens University of Economics and Business, Department of Informatics; Koutsopoulos, Iordanis; Κουτσόπουλος, Ιορδάνης
The appearance of social networking brought major changes in telecommunications and altered the way we perceive the notion of communication. This emerging presence of social networks has shown that network operators cannot remain static by simply providing the conduit for information, but should also encompass the social-networking features that arise. Considering social networks are becoming a norm, it is vital to include them in the investigation of network performance issues. In this work, we consider the problem of using the social networking graph structure as a complement to the upcoming device-to-device (D2D) communication network structure, to the benefit of a wireless operator. Content needs to be sent from a source to a destination. We model the probability that a node accepts to forward the content across a path as a function of its social ties with other nodes. The operator is interested in finding minimum-cost paths that satisfy a minimum probability of delivery guarantee. We formulate the problem as a constrained minimum-cost one and we solve it using a Lagrange relaxation algorithm. The evaluation of the proposed scheme is driven by using a real dataset.
Tackling the cold start problem in product recommendation systems
(12/03/2021) Nikas, Athanasios; Νίκας, Αθανάσιος; Athens University of Economics and Business, Department of Informatics; Vassalos, Vasilios; Koutroumbas, Konstantinos; Koutsopoulos, Iordanis
The need for personalized recommendation has increased hand in hand with the increasing volume of available information through the internet. The overabundance of information, such as commercial products or potential connections in social networks, has driven the development of recommender systems. A recommender system typically learns the user interests from past recorded activities, and consistently processes item candidates to generate user-specificrecommendations. This M.Sc. thesis focuses on the “cold-start” problem in recommendation i.e.,the situation where either a new user desires recommendations, or a brand-new item is to berecommended. Cold-start is especially challenging to recommender systems, since no past interaction data is available to infer the user interests or the item peculiarities. In this thesis two methods are proposed to tackle the cold-start problem, both utilizing available user and item metadata information. The first method combines pre-trained user and item latent factors with metadata features and introduces a modification in the learning process, that enables the model to make recommendations even when latent factors are not available. The second method is a hybrid matrix factorization model that learns user and item representations as combinations of their meta-dataand preference latent factors. The two methods are implemented and evaluated in a banking use case, where banking products are recommended to new customers. The use case dataset is verysparse from the customer’s perspective and has product popularity bias, making therecommendation problem very challenging. Throughout the experimental study, several conclusions are drawn regarding the methods’ performance and parameter sensitivity. Both methods manage to effectively handle the cold-start problem, yielding significant improvements, in terms of Mean Average Precision and Recall, over a baseline popular product recommender and a simple Latent Factor Model in both the warm and cold scenarios.