Solutions descriptions and advancements. OptSFC is designed as an offline training tool that generates an optimized policy of a security function involving the management and orchestration of virtualized network resources. The RL model resulting from the offline training allows deploying the security function in the 5G network with autonomous and efficient prevention and mitigation actions.
Solutions descriptions and advancements. AI and ML contributions Mouseworld This lab has been conceived as a network Digital Twin infrastructure created to produce datasets related to network threats and train ML models. Complex network emulation and testing can be carried out thanks to the equipment and tools. One of the main characteristics of the lab is the ability to create/destroy several network scenarios to run different tests. The virtual scenarios are isolated from each other thanks to the NFV/SDN architecture, and therefore the traffic can be captured using VNF probes. This consistent environment allows evaluating different mitigation tools or versions under the same circumstances, making it perfect for developing and tuning machine learning network instruments.
Solutions descriptions and advancements. AI and ML contributions The targeted poisoning attack on FL system is described as follows. Adversary controls C poisonous nodes where the system has total number of n nodes. We exploit the weaknesses of the existing FoolsGold design and perform a targeted model poisoning attack to circumvent its defence. For the experiments, we use a single layer fully-connected softmax for classification with MNIST digit classifier dataset [74]. FoolsGold algorithm (Figure 22) primarily depends on the Cosine Similarity (CS) of the received gradient updates. We intelligently introduce random noise to the gradient updates at poisonous nodes. The poisonous nodes train their local models with label flipped data and then add the noise before sending them to the central server. A noise vector specifically replaces a part of a given poisonous gradient update to add extra dissimilarity. Only the gradient updates corresponding to the less important features of the model is placed with noise while the vital features required for model training will not be propagated with noise. This ensures that the noise does not reduce the effectiveness of the attack. We use a subset of data at the devices using Stochastic Gradient Descent (SGD) method which reduces the training times. This is done in order to avoid the inconsistencies among training times due to the different amounts of data.
Solutions descriptions and advancements. To fill the aforementioned gaps, we leverage the potential of Deep Learning to build three solutions for tackling the stealthy DDoS attacks in a 5G environment, considering two settings (i.e., network slicing and multi-domain/multi-tenancy) and performing malicious behaviour detection either by using data collected from network layer (i.e., characteristics of network flows) or application layer (i.e., resource usage and service performance metrics).
Solutions descriptions and advancements. To date, there have been very few empirically published accounts of an effective GPS spoofing detection approach that accommodates resource, cost, and environmental constraints. For this purpose, we propose a new solution to devise an effective cellular-enabled UE GPS spoofing detection system, where MLP is used to analyse the statistical features of path losses between UE and base stations (BSs). The proposed approach consists of three processes, namely data sampling, statistical analysis, and MLP prediction. • Data sampling: There are two kinds of data that need to be collected in the processing, namely the actual path losses data and the theoretical path losses data. The actual path losses data are the observations of base stations, and the corresponding theoretical path losses data are computed according to both UE’ GPS positions and BSs’ GPS positions. One of the theoretical path loss models between BS and UE is defined in 3GPP document in [64]. The theoretical path loss model in [64] is adopted in our solution. • Statistical analysis: A statistical analysis is then performed using moments (e.g., Mean Variance Skewness Kurtosis (MVSK)), quartile (e.g., BOX), and probability distributions difference (e.g., Xxxxxxxxxxx Distance (WD)) methods to extract the statistical characteristics of actual and theoretical path losses data. To obtain the metrics of path loss, we consider the number of data transmissions is O in a time duration t. The corresponding actual path loss set 𝐿𝑖𝑣 (𝑡) is expressed as: {𝐿1 (𝑡), … , 𝐿𝑜 (𝑡), … , 𝐿𝑂 (𝑡)} 𝑖𝑣 𝑖𝑣 where 𝐿𝑜 (𝑡) donates the 𝑜𝑡ℎ path loss in t between BS 𝑖 and UE 𝑣. Similarly, the theoretical path loss set is defined as: {𝐿̅1 (𝑡), … , 𝐿̅𝑜 (𝑡), … , 𝐿̅𝑂 (𝑡)} 𝑖𝑣 𝑖𝑣 where 𝐿̅𝑜 (𝑡) denotes the 𝑜𝑡ℎ path loss in t. According to the MVSK method, we have XX (𝑡) = 1 ∑𝑂 𝐿𝑜 (𝑡), 𝑖𝑣 𝑂 𝑜=1 𝑖𝑣 𝐿𝑉 (𝑡) = 1 ∑𝑂 (𝐿𝑜 (𝑡) − XX (𝑡))2, 𝑖𝑣 𝑂 𝑜=1 𝑖𝑣 𝑖𝑣 3 𝐿 𝑆 𝑖𝑣 (𝑡) = 1 𝑂 𝑂 ∑ 𝑜=1 𝐿𝑜 (𝑡)− XX(𝑡) ( 𝑖𝑣 𝑖𝑣 ) , 𝑖𝑣 √𝐿𝑉 (𝑡) 𝐾 1 ∑ 𝐿𝑜 (𝑡)− XX(𝑡) 𝐿𝑖𝑣 (𝑡) = 𝑂 𝑂 𝑜=1 ( 𝑖𝑣 𝑖𝑣 ) , 𝑖𝑣 √𝐿𝑉 (𝑡) where XX (𝑡), 𝐿𝑉 (𝑡), 𝐿𝑆 (𝑡) and ��𝐾 (𝑡) represent, respectively, the mean, variance, skewness, 𝑖𝑣 𝑖𝑣 and kurtosis values of actual path losses in time interval t. Similarly, we can obtain the theoretical ones X̅X (𝑡), 𝐿̅𝑉 (𝑡), 𝐿̅𝑆 (𝑡) and 𝐿̅𝐾 (𝑡). The details on BOX and WD metrics can be found at [65]. 𝑖𝑣 𝑖𝑣 • MLP prediction: The statistical...
Solutions descriptions and advancements. To fill the aforementioned gaps, we leverage the potential of DL to build three solutions for tackling the stealthy DDoS attacks in a 5G environment, considering two settings (i.e., network slicing and multi- domain/multi-tenancy) and performing malicious behaviour detection either by using data collected from network layer (i.e., characteristics of network flows) or application layer (i.e., resource usage and service performance metrics).
Solutions descriptions and advancements. To date, there have been very few empirically published accounts of an effective GPS spoofing detection approach that accommodates resource, cost, and environmental constraints. For this purpose, we propose a new solution to devise an effective cellular-enabled UE GPS spoofing detection system, where MLP is used to analyse the statistical features of path losses between UE and Base Stations (BSs). The proposed approach consists of three processes, namely data sampling, statistical analysis, and MLP prediction. x Data sampling: There are two kinds of data that need to be collected in the processing, namely the actual path losses data and the theoretical path losses data. The actual path losses data are the observations of base stations, and the corresponding theoretical path losses data are ĐŽŵƉƵƚĞĚ ĂĐĐŽƌĚŝŶŐ ƚŽ ďŽƚŚ h ͛ 'W^ ƉŽƐŝƚŝŽŶ path loss models between BS and UE is defined in 3GPP document in [64]. The theoretical path loss model in [64] is adopted in our solution. x Statistical analysis: A statistical analysis is then performed using moments (e.g., Mean Variance Skewness Kurtosis (MVSK)), quartile (e.g., BOX), and probability distributions difference (e.g., Xxxxxxxxxxx Distance (WD)) methods to extract the statistical characteristics of actual and theoretical path losses data. To obtain the metrics of path loss, we consider the number of data transmissions is O in a time duration t. The corresponding actual path loss set ܮ௩ሺݐሻis expressed as: ൛ܮଵሺݐሻǡǥǡܮሺݐሻǡǥǡܮைሺݐሻൟ ௩ ௩ ௩ ௩ where ܮሺݐሻdonates the ௧path loss in t between BS ݅and UE ݒ. Similarly, the theoretical path loss set is defined as: ൛ܮതଵ ሺݐሻǡǥǡܮതሺݐሻǡǥǡܮതைሺݐሻൟ ௩ ௩ ௩ ௩ where ܮതሺݐሻdenotes the ௧path loss in t. According to the MVSK method, we have ܮெሺݐሻ= ଵσை ܮሺݐሻ, ௩ ை ୀଵ ௩ ܮሺݐሻ= ଵσை ሺܮሺݐሻെ ܮெሺݐሻଶ, ௩ ை ୀଵ ௩ ௩ ଷ ௌ ଵ ሺ௧ሻି ಾሺ௧ሻ ܮ ሺݐሻ= σை ቌ ೡ ೡ ቍ , ௩ ை ଵ ୀଵ ೡ ටೇሺ௧ሻ ସ ሺ௧ሻି ಾሺ௧ሻ ܮ ሺݐሻ= σை ቌ ೡ ೡ ቍ , ௩ ை ୀଵ ೡ ටೇሺ௧ሻ where ܮெሺݐሻǡܮሺݐሻǡܮௌሺݐሻand ܮሺݐሻrepresent, respectively, the mean, variance, skewness, ௩ ௩ ௩ ௩ and kurtosis values of actual path losses in time interval t. Similarly, we can obtain the theoretical ones ܮതெሺݐሻǡܮതሺݐሻǡܮതௌሺݐሻand ܮതሺݐሻ. The details on BOX and WD metrics can be found at [65]. ௩ ௩ ௩ ௩ ௩ x MLP prediction: The statistical properties are used as inputs to the devised MLP models (See Figure 17). The input οܮ௫ denotes the ݉௧difference of path losses reported by BS ݅under each of the statistical methods (i.e., MVSK, BOX, WD). ݉௫אሼܯǡܸǡ...
