Web phishing is a form of cybercrime aimed at tricking people into visiting malicious URLs to exfiltrate sensitive data. Since the structure of a malicious URL evolves over time, phishing detection mechanisms that can adapt to such variations are paramount. Furthermore, web phishing detection is an unbalanced classification task, as legitimate URLs outnumber malicious ones in real-life cases. Deep learning (DL) has emerged as a promising technique to minimize concept drift to enhance web phishing detection. Deep reinforcement learning (DRL) combines DL with reinforcement learning (RL); that is, a sequential decision-making paradigm in which the problem to be addressed is expressed as a Markov decision process (MDP). Recent studies have proposed an ad hoc MDP formulation to tackle unbalanced classification tasks called the imbalanced classification Markov decision process (ICMDP). In this paper, we exploit the ICMDP to present a double deep Q-Network (DDQN)-based classifier to address the unbalanced web phishing classification problem. The proposed algorithm is evaluated on a Mendeley web phishing dataset, from which three different data imbalance scenarios are generated. Despite a significant training time, it results in better geometric mean, index of balanced accuracy, F1 score, and area under the ROC curve than other DL-based classifiers combined with data-level sampling techniques in all test cases.

Introduction: Government agencies are now encouraging industries to enhance their security systems to detect and respond proactively to cybersecurity incidents. Consequently, equipping with a security operation center that combines the analytical capabilities of human experts with systems based on Machine Learning (ML) plays a critical role. In this setting, Security Information and Event Management (SIEM) platforms can effectively handle network-related events to trigger cybersecurity alerts. Furthermore, a SIEM may include a User and Entity Behavior Analytics (UEBA) engine that examines the behavior of both users and devices, or entities, within a corporate network.Methods: In recent literature, several contributions have employed ML algorithms for UEBA, especially those based on the unsupervised learning paradigm, because anomalous behaviors are usually not known in advance. However, to shorten the gap between research advances and practice, it is necessary to comprehensively analyze the effectiveness of these methodologies. This paper proposes a thorough investigation of traditional and emerging clustering algorithms for UEBA, considering multiple application contexts, i.e., different user-entity interaction scenarios.Results and discussion: Our study involves three datasets sourced from the existing literature and fifteen clustering algorithms. Among the compared techniques, HDBSCAN and DenMune showed promising performance on the state-of-the-art CERT behavior-related dataset, producing groups with a density very close to the number of users.

Recognition of known malicious patterns through signature-based systems is unsuccessful against malware for which no known signature exists to identify them. These include not only zero-day but also known malicious software able to self-replicate rewriting its own code leaving unaffected its execution, namely metamorphic malware. YARA is a popular malware analysis tool that uses the so-called YARA-rules, which are built to match malicious contents within files or network packets analyzed by an Anti-Virus engine. Sometimes such content is expressed in the form of a byte-signature, i.e., a sequence of operational machine-level code. However, these can be bypassed since malware obfuscation techniques can change these sequences, rewriting them in several equivalent forms. This paper presents YAMME, a YARA-byte-signatures Metamorphic Mutation Engine to strengthen rules against some malware obfuscation techniques deployed in metamorphic mutation engines. First, it rewrites YARA-bye-signatures in several equivalent ways, as a metamorphic mutation engine would do. Second, an optimization phase exploits the YARA-rules syntax constructs to provide several rules formats, making them suitable for different real-world application requirements. YAMME rules have been evaluated on MWOR, G2, NGVCK, and MetaNG datasets, resulting in a better detection rate than that achieved by YARA-rules generated through AutoYara. Furthermore, an analysis of computational overhead required by different YAMME rules formats validates the low impact introduced by the mutation engine at the YARA-rules level.

Packet classification activity performed by a FireWall (FW) introduces high latency in network communications due to the computation time required to check whether any packet matches one of the FW rules. Such a classification process is done by sequentially checking the list of rules until a match is found or the end of the list is reached. Given the complexity of FW rules in some environments, this latency could become relevant. This problem is addressed by ordering the list of FW rules to minimize the classification latency, where the rules with higher activation frequencies are placed accordingly starting from the top of the list. This is not always feasible because dependency constraints between rules could exist: swapping the positions of dependent rules results in a loss of the integrity of the implemented security policy. For this reason, the FW rule ordering problem belongs to the realm of constrained combinatorial optimization. This paper proposes a two-stage algorithm to address this problem. The first stage performs an innovative topological sorting algorithm aimed at finding an optimal ordering for the constrained rules, taking into account the fact that rule activation frequencies are influenced by inter-packet arrival time, which typically obeys Zipf's law. The second stage employs a genetic algorithm to find the optimal ordering of all rules within the list. The proposed approach is evaluated using different filtering lists of different complexity provided by ClassBench. A comparison with other state-of-the-art algorithms addressing the same problem is performed. Furthermore, the performance analysis is extended employing an exact optimization method. The results obtained show the effectiveness of the proposed algorithm in minimizing packet classification latency, while a short reordering time is required.

Data theft through web applications that emulate legitimate platforms constitutes a major network security issue. Countermeasures using artificial intelligence (AI)-based systems are often applied because they can effectively detect malicious websites, which are extremely outnumbered by legitimate ones. In this domain, deep reinforcement learning (DRL) emerges as an attractive field for the development of network intrusion detection models, even in the case of highly skewed class distributions. However, DRL requires training time that increases with data complexity. This paper combines a DRL-based classifier with state-of-the-art feature selection techniques to speed up training while retaining or even improving classification performance. Our experiments used the Mendeley dataset and five different statistical and correlation-based feature-ranking strategies. The results indicated that the selection technique based on the calculation of the Gini index reduces the number of columns in the dataset by 27%, saving more than 10% of training time and significantly improving classification scores compared with the case without selection strategies.