arXiv subjects

Domenico Siracusa

Publications and source records attributed to Domenico Siracusa.

At least 19 recordsLinked to original sources

arXiv2026

Characterizing the Scalability and Performance of Large-Scale AI Training Under Multi-Tenancy

Characterising AI workload performance on modern HPC systems requires understanding both their scalability in isolation and their behaviour under concurrent execution. However, the interplay among parallelisation strategies, network congestion, compute capability, and interconnect technologies remains poorly understood. This work investigates the performance and scalability of AI models up to 2400 GPUs. We quantify the communication overheads and their impact across different interconnects by evaluating scale-up, scale-out, and rack-scale configurations under multiple allocation schemes. Finally, we study how multiple concurrent training jobs interfere with each other by designing a realistic noise model. We design a benchmark suite of AI models to evaluate the performance of five distinct parallelisation strategies across different supercomputing clusters, including Alps, Leonardo, LUMI, JUPITER, NVL72 GB300, and DGX A100. Our work provides a systematic characterization of the scalability and execution efficiency of distributed AI training, while offering key insights into performance behavior under realistic multi-tenant scenarios.

cs.DC↗

arXiv2026

Thou Shall Not Pass: Gatekeeping Outbound TLS Connections

Despite the widespread use of Transport Layer Security (TLS), its security guarantees are frequently compromised by outdated versions and misconfigurations. To analyze this problem, we collected more than 50 million TLS handshakes over a two-week period at our research institution, Fondazione Bruno Kessler, and analyzed three server-selected parameters against the recommendations of four TLS guidelines. Our analysis shows that while the use of insecure or outdated options is minimal, it remains persistent. More importantly, servers are adopting the latest TLS advancements much faster than official guidelines can be updated to provide directives for them. These findings, combined with the difficulty of configuring TLS clients due to their ephemeral, ubiquitous and server-dependent nature, leave users vulnerable to non-standard or outright insecure connections. To address this, we present TLSGatekeeper, a real-time, network-based tool that transparently monitors handshakes, analyzes server parameters, and, based on organizational policy, reports non-compliant connections without requiring client-side modifications. Unlike Next-Generation Firewalls, TLSGatekeeper preserves end-to-end privacy by validating only handshakes, and offers greater flexibility in defining undesired configurations. Our evaluation shows that TLSGatekeeper sustains traffic rates of up to 100 Gbps while preventing insecure connections, with an average added processing delay of 671 ns (TLS 1.3) and 795 ns (TLS 1.2) per handshake packet, making enforcement feasible at scale.

cs.CR↗

arXiv2025

Sharpening Kubernetes Audit Logs with Context Awareness

Kubernetes has emerged as the de facto orchestrator of microservices, providing scalability and extensibility to a highly dynamic environment. It builds an intricate and deeply connected system that requires extensive monitoring capabilities to be properly managed. To this account, K8s natively offers audit logs, a powerful feature for tracking API interactions in the cluster. Audit logs provide a detailed and chronological record of all activities in the system. Unfortunately, K8s auditing suffers from several practical limitations: it generates large volumes of data continuously, as all components within the cluster interact and respond to user actions. Moreover, each action can trigger a cascade of secondary events dispersed across the log, with little to no explicit linkage, making it difficult to reconstruct the context behind user-initiated operations. In this paper, we introduce K8NTEXT, a novel approach for streamlining K8s audit logs by reconstructing contexts, i.e., grouping actions performed by actors on the cluster with the subsequent events these actions cause. Correlated API calls are automatically identified, labeled, and consistently grouped using a combination of inference rules and a Machine Learning model, largely simplifying data consumption. We evaluate K8NTEXT's performance, scalability, and expressiveness both in systematic tests and with a series of use cases. We show that it consistently provides accurate context reconstruction, even for complex operations involving 50, 100 or more correlated actions, achieving over 95 percent accuracy across the entire spectrum, from simple to highly composite actions.

cs.CR↗

arXiv2025

Disruption-aware Microservice Re-orchestration for Cost-efficient Multi-cloud Deployments

Multi-cloud environments enable a cost-efficient scaling of cloud-native applications across geographically distributed virtual nodes with different pricing models. In this context, the resource fragmentation caused by frequent changes in the resource demands of deployed microservices, along with the allocation or termination of new and existing microservices, increases the deployment cost. Therefore, re-orchestrating deployed microservices on a cheaper configuration of multi-cloud nodes offers a practical solution to restore the cost efficiency of deployment. However, the rescheduling procedure causes frequent service interruptions due to the continuous termination and rebooting of the containerized microservices. Moreover, it may potentially interfere with and delay other deployment operations, compromising the stability of the running applications. To address this issue, we formulate a multi-objective integer linear programming (ILP) problem that computes a microservice rescheduling solution capable of providing minimum deployment cost without significantly affecting the service continuity. At the same time, the proposed formulation also preserves the quality of service (QoS) requirements, including latency, expressed through microservice co-location constraints. Additionally, we present a heuristic algorithm to approximate the optimal solution, striking a balance between cost reduction and service disruption mitigation. We integrate the proposed approach as a custom plugin of the Kubernetes (K8s) scheduler. Results reveal that our approach significantly reduces multi-cloud deployment costs and service disruptions compared to the benchmark schemes, while ensuring QoS requirements are consistently met.

cs.NI↗

arXiv2024

INTELLECT: Adapting Cyber Threat Detection to Heterogeneous Computing Environments

The widespread adoption of cloud computing, edge, and IoT has increased the attack surface for cyber threats. This is due to the large-scale deployment of often unsecured, heterogeneous devices with varying hardware and software configurations. The diversity of these devices attracts a wide array of potential attack methods, making it challenging for individual organizations to have comprehensive knowledge of all possible threats. In this context, powerful anomaly detection models can be developed by combining data from different parties using Federated Learning. FL enables the collaborative development of ML-based IDSs without requiring the parties to disclose sensitive training data, such as network traffic or sensor readings. However, deploying the resulting models can be challenging, as they may require more computational resources than those available on target devices with limited capacity or already allocated for other operations. Training device-specific models is not feasible for an organization because a significant portion of the training data is private to other participants in the FL process. To address these challenges, this paper introduces INTELLECT, a novel solution that integrates feature selection, model pruning, and fine-tuning techniques into a cohesive pipeline for the dynamic adaptation of pre-trained ML models and configurations for IDSs. Through empirical evaluation, we analyze the benefits of INTELLECT's approach in tailoring ML models to the specific resource constraints of an organization's devices and measure variations in traffic classification accuracy resulting from feature selection, pruning, and fine-tuning operations. Additionally, we demonstrate the advantages of incorporating knowledge distillation techniques while fine-tuning, enabling the ML model to consistently adapt to local network patterns while preserving historical knowledge.

cs.CR↗

arXiv2024

A Proactive Decoy Selection Scheme for Cyber Deception using MITRE ATT&CK

Cyber deception allows compensating the late response of defenders countermeasures to the ever evolving tactics, techniques, and procedures (TTPs) of attackers. This proactive defense strategy employs decoys resembling legitimate system components to lure stealthy attackers within the defender environment, slowing and/or denying the accomplishment of their goals. In this regard, the selection of decoys that can expose the techniques used by malicious users plays a central role to incentivize their engagement. However, this is a difficult task to achieve in practice, since it requires an accurate and realistic modeling of the attacker capabilities and his possible targets. In this work, we tackle this challenge and we design a decoy selection scheme that is supported by an adversarial modeling based on empirical observation of real-world attackers. We take advantage of a domain-specific threat modelling language using MITRE ATT&CK framework as source of attacker TTPs targeting enterprise systems. In detail, we extract the information about the execution preconditions of each technique as well as its possible effects on the environment to generate attack graphs modeling the adversary capabilities. Based on this, we formulate a graph partition problem that minimizes the number of decoys detecting a corresponding number of techniques employed in various attack paths directed to specific targets. We compare our optimization-based decoy selection approach against several benchmark schemes that ignore the preconditions between the various attack steps. Results reveal that the proposed scheme provides the highest interception rate of attack paths using the lowest amount of decoys.

cs.CR↗

arXiv2024

Exploiting Kubernetes' Image Pull Implementation to Deny Node Availability

Kubernetes (K8s) has grown in popularity over the past few years to become the de-facto standard for container orchestration in cloud-native environments. While research is not new to topics such as containerization and access control security, the Application Programming Interface (API) interactions between K8s and its runtime interfaces have not been studied thoroughly. In particular, the CRI-API is responsible for abstracting the container runtime, managing the creation and lifecycle of containers along with the downloads of the respective images. However, this decoupling of concerns and the abstraction of the container runtime renders K8s unaware of the status of the downloading process of the container images, obstructing the monitoring of the resources allocated to such process. In this paper, we discuss how this lack of status information can be exploited as a Denial of Service attack in a K8s cluster. We show how such attacks can impact worker nodes, generating up to 95% average CPU usage, prevent downloads of new container images, and increase I/O and network usage for a potentially unlimited amount of time. We argue that solving this problem would require a radical architectural change in the relationship between K8s and the CRI-API, which would be unfeasible in the short term. Thus, as a stopgap solution, we propose MAGI: an eBPF-based, proof-of-concept mitigation that detects and terminates potential attacks.

cs.CR↗

arXiv2024

Cost Minimization in Multi-cloud Systems with Runtime Microservice Re-orchestration

Multi-cloud systems facilitate a cost-efficient and geographically-distributed deployment of microservice-based applications by temporary leasing virtual nodes with diverse pricing models. To preserve the cost-efficiency of multi-cloud deployments, it is essential to redeploy microservices onto the available nodes according to a dynamic resource configuration, which is often performed to better accommodate workload variations. However, this approach leads to frequent service disruption since applications are continuously shutdown and redeployed in order to apply the new resource assignment. To overcome this issue, we propose a re-orchestration scheme that migrates microservice at runtime based on a rolling update scheduling logic. Specifically, we propose an integer linear optimization problem that minimizes the cost associated to multi-cloud virtual nodes and that ensures that delay-sensitive microservices are co-located on the same regional cluster. The resulting rescheduling order guarantees no service disruption by repacking microservices between the available nodes without the need to turn off the outdated microservice instance before redeploying the updated version. In addition, we propose a two-step heuristic scheme that effectively approximates the optimal solution at the expense of close-to-zero service disruption and QoS violation probability. Results show that proposed schemes achieve better performance in terms of cost mitigation, low service disruption and low QoS violation probability compared to baseline schemes replicating Kubernetes scheduler functionalities.

cs.NI↗

arXiv2023

Introducing Packet-Level Analysis in Programmable Data Planes to Advance Network Intrusion Detection

Programmable data planes offer precise control over the low-level processing steps applied to network packets, serving as a valuable tool for analysing malicious flows in the field of intrusion detection. Albeit with limitations on physical resources and capabilities, they allow for the efficient extraction of detailed traffic information, which can then be utilised by Machine Learning (ML) algorithms responsible for identifying security threats. In addressing resource constraints, existing solutions in the literature rely on compressing network data through the collection of statistical traffic features in the data plane. While this compression saves memory resources in switches and minimises the burden on the control channel between the data and the control plane, it also results in a loss of information available to the Network Intrusion Detection System (NIDS), limiting access to packet payload, categorical features, and the semantic understanding of network communications, such as the behaviour of packets within traffic flows. This paper proposes P4DDLe, a framework that exploits the flexibility of P4-based programmable data planes for packet-level feature extraction and pre-processing. P4DDLe leverages the programmable data plane to extract raw packet features from the network traffic, categorical features included, and to organise them in a way that the semantics of traffic flows are preserved. To minimise memory and control channel overheads, P4DDLe selectively processes and filters packet-level data, so that only the features required by the NIDS are collected. The experimental evaluation with recent Distributed Denial of Service (DDoS) attack data demonstrates that the proposed approach is very efficient in collecting compact and high-quality representations of network flows, ensuring precise detection of DDoS attacks.

cs.CR↗

arXiv2023

Resource-aware Cyber Deception for Microservice-based Applications

Cyber deception can be a valuable addition to traditional cyber defense mechanisms, especially for modern cloud-native environments with a fading security perimeter. However, pre-built decoys used in classical computer networks are not effective in detecting and mitigating malicious actors due to their inability to blend with the variety of applications in such environments. On the other hand, decoys cloning the deployed microservices of an application can offer a high-fidelity deception mechanism to intercept ongoing attacks within production environments. However, to fully benefit from this approach, it is essential to use a limited amount of decoy resources and devise a suitable cloning strategy to minimize the impact on legitimate services performance. Following this observation, we formulate a non-linear integer optimization problem that maximizes the number of attack paths intercepted by the allocated decoys within a fixed resource budget. Attack paths represent the attacker's movements within the infrastructure as a sequence of violated microservices. We also design a heuristic decoy placement algorithm to approximate the optimal solution and overcome the computational complexity of the proposed formulation. We evaluate the performance of the optimal and heuristic solutions against other schemes that use local vulnerability metrics to select which microservices to clone as decoys. Our results show that the proposed allocation strategy achieves a higher number of intercepted attack paths compared to these schemes while requiring approximately the same number of decoys.

cs.CR↗

arXiv2022

FLAD: Adaptive Federated Learning for DDoS Attack Detection

Federated Learning (FL) has been recently receiving increasing consideration from the cybersecurity community as a way to collaboratively train deep learning models with distributed profiles of cyber threats, with no disclosure of training data. Nevertheless, the adoption of FL in cybersecurity is still in its infancy, and a range of practical aspects have not been properly addressed yet. Indeed, the Federated Averaging algorithm at the core of the FL concept requires the availability of test data to control the FL process. Although this might be feasible in some domains, test network traffic of newly discovered attacks cannot be always shared without disclosing sensitive information. In this paper, we address the convergence of the FL process in dynamic cybersecurity scenarios, where the trained model must be frequently updated with new recent attack profiles to empower all members of the federation with the latest detection features. To this aim, we propose FLAD (adaptive Federated Learning Approach to DDoS attack detection), an FL solution for cybersecurity applications based on an adaptive mechanism that orchestrates the FL process by dynamically assigning more computation to those members whose attacks profiles are harder to learn, without the need of sharing any test data to monitor the performance of the trained model. Using a recent dataset of DDoS attacks, we demonstrate that FLAD outperforms state-of-the-art FL algorithms in terms of convergence time and accuracy across a range of unbalanced datasets of heterogeneous DDoS attacks. We also show the robustness of our approach in a realistic scenario, where we retrain the deep learning model multiple times to introduce the profiles of new attacks on a pre-trained model.

cs.CR↗

arXiv2022

GADoT: GAN-based Adversarial Training for Robust DDoS Attack Detection

Machine Learning (ML) has proven to be effective in many application domains. However, ML methods can be vulnerable to adversarial attacks, in which an attacker tries to fool the classification/prediction mechanism by crafting the input data. In the case of ML-based Network Intrusion Detection Systems (NIDSs), the attacker might use their knowledge of the intrusion detection logic to generate malicious traffic that remains undetected. One way to solve this issue is to adopt adversarial training, in which the training set is augmented with adversarial traffic samples. This paper presents an adversarial training approach called GADoT, which leverages a Generative Adversarial Network (GAN) to generate adversarial DDoS samples for training. We show that a state-of-the-art NIDS with high accuracy on popular datasets can experience more than 60% undetected malicious flows under adversarial attacks. We then demonstrate how this score drops to 1.8% or less after adversarial training using GADoT.

cs.CR↗

arXiv2021

A Cost-Effective Workload Allocation Strategy for Cloud-Native Edge Services

Nowadays IoT applications consist of a collection of loosely coupled modules, namely microservices, that can be managed and placed in a heterogeneous environment consisting of private and public resources. It follows that distributing the application logic introduces new challenges in guaranteeing performance and reducing costs. However, most existing solutions are focused on reducing pay-per-use costs without considering a microservice-based architecture. We propose a cost-effective workload allocation for microservice-based applications. We model the problem as an integer programming problem and we formulate an efficient and near-optimal heuristic solution given the NP-hardness of the original problem. Numerical results demonstrate the good performance of the proposed heuristic in terms of cost reduction and performance with respect to optimal and state-of-the-art solutions. Moreover, an evaluation conducted in a Kubernetes cluster running in an OpenStack ecosystem confirms the feasibility and the validity of the proposed solution.

cs.NI↗

arXiv2021

In-Network Volumetric DDoS Victim Identification Using Programmable Commodity Switches

Volumetric distributed Denial-of-Service (DDoS) attacks have become one of the most significant threats to modern telecommunication networks. However, most existing defense systems require that detection software operates from a centralized monitoring collector, leading to increased traffic load and delayed response. The recent advent of Data Plane Programmability (DPP) enables an alternative solution: threshold-based volumetric DDoS detection can be performed directly in programmable switches to skim only potentially hazardous traffic, to be analyzed in depth at the controller. In this paper, we first introduce the BACON data structure based on sketches, to estimate per-destination flow cardinality, and theoretically analyze it. Then we employ it in a simple in-network DDoS victim identification strategy, INDDoS, to detect the destination IPs for which the number of incoming connections exceeds a pre-defined threshold. We describe its hardware implementation on a Tofino-based programmable switch using the domain-specific P4 language, proving that some limitations imposed by real hardware to safeguard processing speed can be overcome to implement relatively complex packet manipulations. Finally, we present some experimental performance measurements, showing that our programmable switch is able to keep processing packets at line-rate while performing volumetric DDoS detection, and also achieves a high F1 score on DDoS victim identification.

cs.NI↗

arXiv2021

Tracking Normalized Network Traffic Entropy to Detect DDoS Attacks in P4

Distributed Denial-of-Service (DDoS) attacks represent a persistent threat to modern telecommunications networks: detecting and counteracting them is still a crucial unresolved challenge for network operators. DDoS attack detection is usually carried out in one or more central nodes that collect significant amounts of monitoring data from networking devices, potentially creating issues related to network overload or delay in detection. The dawn of programmable data planes in Software-Defined Networks can help mitigate this issue, opening the door to the detection of DDoS attacks directly in the data plane of the switches. However, the most widely-adopted data plane programming language, namely P4, lacks supporting many arithmetic operations, therefore, some of the advanced network monitoring functionalities needed for DDoS detection cannot be straightforwardly implemented in P4. This work overcomes such a limitation and presents two novel strategies for flow cardinality and for normalized network traffic entropy estimation that only use P4-supported operations and guarantee a low relative error. Additionally, based on these contributions, we propose a DDoS detection strategy relying on variations of the normalized network traffic entropy. Results show that it has comparable or higher detection accuracy than state-of-the-art solutions, yet being simpler and entirely executed in the data plane.

cs.NI↗

arXiv2020

DAICS: A Deep Learning Solution for Anomaly Detection in Industrial Control Systems

Deep Learning is emerging as an effective technique to detect sophisticated cyber-attacks targeting Industrial Control Systems (ICSs). The conventional approach to detection in literature is to learn the "normal" behaviour of the system, to be then able to label noteworthy deviations from it as anomalies. However, during operations, ICSs inevitably and continuously evolve their behaviour, due to e.g., replacement of devices, workflow modifications, or other reasons. As a consequence, the accuracy of the anomaly detection process may be dramatically affected with a considerable amount of false alarms being generated. This paper presents DAICS, a novel deep learning framework with a modular design to fit in large ICSs. The key component of the framework is a 2-branch neural network that learns the changes in the ICS behaviour with a small number of data samples and a few gradient updates. This is supported by an automatic tuning mechanism of the detection threshold that takes into account the changes in the prediction error under normal operating conditions. In this regard, no specialised human intervention is needed to update the other parameters of the system. DAICS has been evaluated using publicly available datasets and shows an increased detection rate and accuracy compared to state of the art approaches, as well as higher robustness to additive noise.

cs.CR↗

arXiv2020

LUCID: A Practical, Lightweight Deep Learning Solution for DDoS Attack Detection

Distributed Denial of Service (DDoS) attacks are one of the most harmful threats in today's Internet, disrupting the availability of essential services. The challenge of DDoS detection is the combination of attack approaches coupled with the volume of live traffic to be analysed. In this paper, we present a practical, lightweight deep learning DDoS detection system called LUCID, which exploits the properties of Convolutional Neural Networks (CNNs) to classify traffic flows as either malicious or benign. We make four main contributions; (1) an innovative application of a CNN to detect DDoS traffic with low processing overhead, (2) a dataset-agnostic preprocessing mechanism to produce traffic observations for online attack detection, (3) an activation analysis to explain LUCID's DDoS classification, and (4) an empirical validation of the solution on a resource-constrained hardware platform. Using the latest datasets, LUCID matches existing state-of-the-art detection accuracy whilst presenting a 40x reduction in processing time, as compared to the state-of-the-art. With our evaluation results, we prove that the proposed approach is suitable for effective DDoS detection in resource-constrained operational environments.

cs.CR↗

arXiv2019

Dynamic and Application-Aware Provisioning of Chained Virtual Security Network Functions

A promising area of application for Network Function Virtualization is in network security, where chains of Virtual Security Network Functions (VSNFs), i.e., security-specific virtual functions such as firewalls or Intrusion Prevention Systems, can be dynamically created and configured to inspect, filter or monitor the network traffic. However, the traffic handled by VSNFs could be sensitive to specific network requirements, such as minimum bandwidth or maximum end-to-end latency. Therefore, the decision on which VSNFs should apply for a given application, where to place them and how to connect them, should take such requirements into consideration. Otherwise, security services could affect the quality of service experienced by customers. In this paper we propose PESS (Progressive Embedding of Security Services), a solution to efficiently deploy chains of virtualised security functions based on the security requirements of individual applications and operators' policies, while optimizing resource utilization. We provide the PESS mathematical model and heuristic solution. Simulation results show that, compared to state-of-the-art application-agnostic VSNF provisioning models, PESS reduces computational resource utilization by up to 50%, in different network scenarios. This result ultimately leads to a higher number of provisioned security services and to up to a 40% reduction in end-to-end latency of application traffic.

cs.NI↗