Problem and challenges. Pervasive Vehicle-to-everything (V2X) connectivity and the emergence of effective data-driven methods based on AI/ML drive a paradigm shift towards Connected and Automated Mobility (CAM) services and applications [18]. A key functionality in vehicular systems that can benefit from AI/ML is security, which is essential for ensuring road safety in CAM environments. V2X security threats and attacks can either originate from malicious outsiders which are vehicles/users exogenous to the original system, or insiders which are already authenticated and possess valid credentials to interact with other legitimate entities in the system. While outsider attacks can be efficiently addressed even in highly dense V2X scenarios with a proper extension of the 5G Authentication and Key Agreement (5G-AKA) procedure, as shown in [19], insider attacks are often difficult to detect and contain, particularly when attackers behave intelligently while conforming to normal system behaviour. For example, an already authenticated vehicle may be able to intentionally transmit false kinematic information (e.g., position, speed, acceleration, heading-angle data) in its broadcast messages and cause disruption in the network. Seemingly abnormal vehicular activity originated from malicious actors (e.g., vehicles) may take the form of highly sophisticated attacks. Such malicious/selfish behaviours from such rogue insiders are commonly referred to as misbehaviours in V2X, and they pose a serious threat when transmitting erroneous/incorrect data in safety-critical situations. Ensuring the semantic correctness of exchanged V2X information is thus of paramount importance.
Problem and challenges. The advent of 5G and its many advances over previous mobile technologies - much lower latency, huge bandwidth, the possibility to connect many more devices per square meter, and so on and so forth - will not just bring benefits. It turns out that all these advances in mobile network performance will provide the perfect breeding ground for attacks. Denial-of-service attacks, in particular, will benefit the most from this: larger bandwidth will allow much more traffic to be sent per device, and the fact that many more devices can be concurrently connected to the network (proliferation of IoT devices) will allow much larger, and much more powerful botnets to be created in order to carry out these types of attacks much more effectively, especially empowering distributed denial-of-service attacks (DDoS) The main challenge that arises from the previous aspects is an effective detection for traditional DDoS attacks (e.g., flooding attacks) and also for more advanced stealthy DDoS attacks (e.g., SlowDoS attacks). For this purpose, we aim to leverage AI techniques, particularly Deep Learning techniques, for an efficient detection and mitigation of such attacks in 5G environments.
Problem and challenges. Many research efforts have been devoted to tackle DDoS attacks leveraging ML and/or SDN. Xxxxx et al. [44] proposed an intelligent method for detecting network-layer DDoS attacks in an SDN environment. The proposed method uses a Self-Organizing Maps (SOM) [45] model, an unsupervised artificial neural network, trained on traffic flow features. The contribution in [46] rely on DNN models to detect intrusion in an SDN network. The authors in [47] devised a ML-based collaborative DDoS mitigation strategy in a multi-SDN controller environment. The detection is performed Naive Bayes classifier based on flow features extracted by the SDN controller. Upon detection of malicious ďĞŚĂǀŝŽƵƌ͕ ƚŚĞ ^ E ĐŽŶƚƌŽůůĞƌ ŝŶ ƚŚĞ ĂƚƚĂĐŬĞƌ͛ based flow. Similar to [40], the work in [42] , [43] consider only network-layer attacks. Moreover, the proposed models are trained on NSL-KDD, a relatively old dataset that cannot reflect the current trend in network attacks. Hong et al.[48] devised an SDN-assisted defence method to detect and mitigate slow HTTP DDoS attacks. The defence solution is deployed as a SDN application and triggered by the web server when the number of open connections that sent incomplete HTTP requests exceeds a given threshold. The major weakness of threshold-based schemes is their lack of accuracy. In fact, threshold-based schemes are unsuitable for detecting application-layer DDoS attacks due to the resemblance between the traffic patterns generated by those attacks and benign activities. The authors in [49] demonstrated the potential of ML techniques in detecting low-rate application-layer DDoS using the characteristics of malicious TCP flows. A detection accuracy of over 97% has been achieved using K-Nearest Neighbour, Decision Trees and DNN techniques. Some solutions related to the detection of DDoS attacks over 5G multi-tenant networks have been presented in recent years. For instance, Mamolar et. al [50] proposed an extension of the well-known Intrusion Detection System (IDS) Snort, capable of detecting DDoS attacks in real time, to support 5G multi-tenant traffic, so it can be deployed in a multi-tenant 5G environment. However, they do not leverage any AI technique, so we consider this approach too static and inappropriate for such dynamic network environments as those found in 5G. Furthermore, very few contributions have focused on addressing the issue in 5G network slicing environment leveraging mainly the resource isolation concept (e.g., [51]). However,...
Problem and challenges. B5G network opens new opportunities to operators to apply ML to solve multiple problems, including advanced security management. To achieve these results, it is needed to invest a non-negligible quantity of effort in the data engineering process, including data sources identification, data transformation, and to evaluate conditions such as frequency and quality of the data. Network range- digital twin (NDT) appears as a potential solution for assessing solutions related to AI/ML architectures. This includes generation, collection, and transformation of data to design and test different ML models in an emulated environment before deploying in production, reducing the cost and investment. This model could fit for offline ML training, and ML inference engine delivery. Figure 18 shows the holistic process combining several enablers. Figure 18: Data to ML cycle Mouseworld acts as the 5G twin environment for network traffic generation and delivery. This network flows related information is collected and aggregated through a data collector. The output generated is delivered to design an ML model. The outcome is integrated into the Smart Traffic Analyzer and validated in the Mouseworld. If the model does not achieve the expectations or the traffic scenarios evolves, a redesign can be done.
Problem and challenges. The standard setting in ML considers centralized datasets which are tightly integrated into the system. However, in most real-world scenarios, data is usually distributed among multiple entities. More specifically, centralized data collection is challenging due to the higher communication cost for sending data, when the devices create large volumes of data, serious privacy issues coming with the sharing of sensitive data, overfitting issues with the small datasets and the biased local datasets. As a solution, federated training is proposed where each user and server collaborate to train a unified neural network model. This ML approach was formally published by Google in 2016 as Federated Learning (FL). Simply, FL is a distributed learning concept, where End devices or workers are participating for learning process. The central entity or parameter server shares the training model and aggregates the local model updates coming from workers. Workers train the shared model locally using their own data and send the trained model back to the central server. Central server aggregates the received models and shares the aggregated model with workers. The final model needs to be as good as the centralized solution (ideally), or at least better than what each party can learn on its own. Typically, FL brings the advantages in terms of improving privacy awareness, low communication overhead, and low latency. Most importantly, FL is suitable to address the distributed networking scenarios in the more complex networks. However, FL is vulnerable to poisoning attacks by design (Figure 21). The central server can be poisoned using minimum of one adversarial worker. This will affect the learning process of the entire network. The problem is that the central server cannot guarantee that the workers provide accurate local models and have no control over the level of security at each worker. Another issue is that it is possible to encounter a single point of failure at the central server. Therefore, it is necessary to implement defence mechanisms at the central server to distinguish between poisonous and honest users. It is challenging since the central server has no validation data for verification of the model updates received by the workers. Figure 21: System model of poisoning attacks for Federated Learning
