In the modern age we live in, the internet has become an essential part of our daily life. A significant portion of our personal data is stored online and organizations run their business online. In addition, with the development of the internet, many devices such as autonomous systems, investment portfolio tools and entertainment tools in our homes and workplaces have become or are becoming intelligent. In parallel with this development, cyberattacks aimed at damaging smart systems are increasing day by day. As cyberattack methods become more sophisticated, the damage done by attackers is increasing exponentially. Traditional computer algorithms may be insufficient against these attacks in the virtual world. Therefore, artificial intelligence-based methods are needed. Reinforcement Learning (RL), a machine learning method, is used in the field of cyber security. Although RL for cyber security is a new topic in the literature, studies are carried out to predict, prevent and stop attacks. In this study; we reviewed the literature on RL's penetration testing, intrusion detection systems (IDS) and cyberattacks in cyber security.

Ağ tabanlı Saldırı Tespit Sistemleri (NIDS), ağda bulunan tüm cihazlardan gelen trafiği izlemek ve analiz etmek için kullanılır. Makine Öğrenimi (ML) tabanlı NIDS, günümüzde bilgisayar ağlarını siber saldırılara karşı korumak için önemli araçlardan biridir. ML tabanlı NIDS'in eğitimi ve değerlendirilmesi için ağ veri özellikleri önemli bir etkiye sahiptir. Bu nedenle ML modelinin doğruluğunu ve performansını değerlendirmek için birden çok veri kümesinin ortak temel özellik kümesi içermesi gerekir. Bu çalışmada ortak NetFlow özelliklerine sahip NIDS veri setleri (NF-UNSW-NB15, NF-BoT-IoT, NF-ToN-IoT ve NF-CSE-CIC-IDS2018) kullanılarak ikili sınıflandırma yapılmıştır. Veri setlerindeki saldırı ve normal akış (saldırı yok) sınıfları dengesiz dağılım göstermektedir. Bunun üstesinden gelmek için Rastgele Alt Örnekleme yöntemi kullanılmıştır. Sınıflandırma yöntemleri olarak Rastgele Orman, K-En Yakın Komşuluk, Destek Vektör Makineleri ve Yapay Sinir Ağları algoritmaları kullanılmıştır. Farklı veri setlerinin yeniden örneklenmiş durumlarına, ML yöntemleri kullanılarak doğruluk ve performansları karşılaştırılmıştır. Bu çalışma kapsamında kullanılmış olan dört veri seti içinde en iyi sonucu Rastgele Orman algoritması vermiştir.

As a direct consequence of the unrelenting march of technological innovation, the use of the Internet has become an unavoidable condition for the life of modern humans. The Internet has increased both the quantity and range of situations in which information products can be useful or non-useful. It’s no surprise that as the number of different systems and users has grown, so have the number of different ways to exploit those systems. A security issue has arisen with such diversity and growth. Its diversity and increase in quantity introduce new system weaknesses and thus new attack strategies. Methods for detecting both internal and external attacks are suggested as a solution to this issue. The purpose of this research, a Convolutional Neural Network was utilized to identify intrusions, also known as attacks for the imbalanced class distribution in the NF-BoT-IoT data set, Synthetic Minority Over Sampling Technique, Random Over Sampling and Random Under Sampling methods were used. K-Fold Cross Validation, one of the strategies for splitting the data set, was utilized to evaluate the performance of classification models and to train the developed model. The model’s performance was evaluated using the accuracy, precision, recall, and F1-score performance criteria.

Money laundering seriously threatens economic stability by legitimizing illegal gains. Despite the transparency and security advantages offered by the blockchain technology, anonymity can create a platform for concealing illegal activities. Therefore, detecting and preventing money laundering activities in blockchain networks are of great importance. This study classifies money transfers during Bitcoin transactions as licit or illicit. By working on the Elliptic dataset to detect money laundering activities in the Bitcoin network, we examined money laundering traffic data using a graph data structure. This study presents a novel method for analyzing complex networks in money laundering as a chaotic time series. First, we increase the number of features of the graph nodes and convert them into a time series. By transferring the obtained time series to the phase space, we calculated the Lyapunov Exponents and aimed to capture the changes and uncertainties in the dynamic structure of the system more accurately using different embedding dimensions. We reconstructed the graph structure representing the transactions based on the feature vectors of these exponents, and classified the transactions using the Graph Convolutional Network method. In our study, we achieved a precision of 92.5%, recall of 92.1%, F1-score of 92.3%, and accuracy of 86.2%. These results demonstrate the effectiveness and reliability of our model in detecting money laundering. This study offers a novel approach for classifying chaotic structures in anti money laundering.

Recent advances in machine learning, particularly with regard to deep learning, help to recognize and classify objects in medical images. In this study, endoscopy images were examined and deep learning method was used to classify healthy and polyp cells. For the proposed system, a database was created from the archives of General Surgery Department Endoscopy Unit in Kutahya Evliya Celebi Training and Research Hospital. The database contains 93 polyps and 216 normal images from 54 archive records. For image multiplexing, a total of 1236 images were obtained by rotating each image 90 degrees around its axis. While 2/3 of the randomly selected data from this obtained data was used for training the model, the rest of the data was reserved for testing. K-fold Cross Validation method was used to reduce the variability of performance results. In this study, 48 different models were created by using different activation and optimization functions to find the best classification model in deep learning. According to the experimental results, it was observed that accuracy of the models depends on the selected parameters; the best model with the accuracy rate of 91% was obtained with 64 neurons in the hidden layer, ReLU activation function and RmsProp optimization method whereas the worst model with the accuracy rate of 76% was obtained with 32 neurons in the hidden layer, Tanh activation and PmsProp optimization functions. Accordingly, classification performance of polyp images can be optimized by utilizing different activation and optimization methods during the design of deep learning models.

The Quantum Approximate Optimization Algorithm (QAOA) reduces to a hard classical subproblem: selecting the variational angles that parameterize the circuit. The landscape is non-convex and dense with poor local optima, and because every evaluation is costly, the budget is the binding resource. We present GRACE, a closed-loop framework for QAOA parameter setting on weighted MaxCut that couples two components under strict, uniform per-evaluation budget accounting: a graph neural network (GNN) that predicts high-quality initial angles from the problem graph in a single circuit evaluation, and a metaheuristic escape operator invoked when the loop stalls to rescue difficult instances. The escape stage is operator-portable: the framework accepts any of a broad operator family and benefits from all of them. We compare eleven escape operators spanning ant colony, differential evolution, swarm, evolutionary, and recent nature-inspired search, plus no-escape and random-restart controls, under an identical per-escape budget and at tuned coefficients. Continuous ant colony optimization is strongest; every metaheuristic beats the no-escape control while random restart does not, showing the benefit stems from structured search rather than perturbation. On a 60-instance, three-family benchmark at depth p=5, GRACE attains a mean approximation ratio of 0.901 and the best rank against eight baselines, superior on all 60 instances. Its defining property is reliability: a per-instance standard deviation (0.016) far below that of the strong INTERP and FOURIER schedules. A depth sweep and a size study show the operator ranking is stable at practical depths and that GRACE's margin widens with problem size.