Traffic Classification (TC) is pivotal for network management, cybersecurity, and Quality of Experience (QoE) monitoring. However, while Deep Learning (DL) has significantly advanced TC, most existing works assume static, idealized conditions, overlooking key challenges of real-world deployments—such as traffic variability, routing asymmetries, out-of-order packet arrivals, and partial visibility at the Vantage Points (VPs). This motivates the need for robustness evaluations under such scenarios. In this work, we investigate the robustness of state-of-the-art (SOTA) TC models under realistic, yet controlled, perturbation scenarios. Specifically, we introduce novel, model-agnostic traffic perturbations—simulating time jitter, retransmissions, and partial visibility—to reflect conditions commonly encountered in live network traffic. We evaluate our approach on three public datasets—i.e., VPN-16 , MIRAGE-19 , and MIRAGE-24 —and show how Mimetic-Enhanced , a multimodal model, tends to outperform two representative single-modal counterparts both in terms of TC effectiveness on clean traffic and robustness under perturbations. Nonetheless, our analysis also reveals that multimodal models remain vulnerable under specific perturbation settings. To address this limitation, we propose a model-agnostic perturbation-aware training framework based on Supervised Data Augmentation ( Aug ) and Contrastive Learning ( CL )—considering both self-supervised and supervised variants. Unlike architecture-specific solutions, our approach operates at the learning strategy level , allowing it to be seamlessly applied to diverse classifiers without requiring structural modifications. Adopting Mimetic-Enhanced as a primary multimodal case study, we integrate the proposed strategies into its two-stage training pipeline. Experimental results demonstrate that perturbation-aware training not only improves TC effectiveness on clean (i.e., unperturbed) traffic—particularly when applied across both training stages—but also significantly strengthens the model’s robustness under diverse and realistic perturbation scenarios. Furthermore, we investigate Out-of-Distribution (OOD) detection, model calibration, and TC effectiveness in low-data regimes. Finally, we explicitly demonstrate the framework’s generalizability by validating it on other SOTA architectures, spanning both single- and multi-modal approaches.

Generative AI (GenAI) chatbots are now pervasive in digital ecosystems, fundamentally reshaping user interactions over the Internet. Their reliance on an always-online, cloud-centric operating model introduces novel traffic dynamics that challenge practical network management. Despite the critical need to anticipate these changes in network demand, the traffic characterization of these chatbots remains largely underexplored. To fill this gap, this study presents an in-depth traffic analysis of ChatGPT , Copilot , and Gemini used via Android mobile apps. Using a dedicated capture architecture, we collect two complementary datasets, combining unconstrained user interactions with a controlled workload of selected prompts for both text and image generation. This dual design allows us to address practical research questions on the distinctiveness of chatbot traffic, its divergence from that of conventional messaging apps, and its novel implications for network usage. To this end, we provide a multi-granular traffic characterization and model packet-sequence dynamics to uncover the underlying transmission mechanisms. Our analysis reveals app-/content-specific traffic patterns and distinctive protocol footprints. We highlight the predominance of TLS, with Gemini extensively leveraging QUIC, ChatGPT exclusively using TLS 1.3, and characteristic Server Name Indication (SNI) values. Through occlusion analysis, we quantify the reliance on SNI for traffic visibility, demonstrating that masking this field reduces classification performance by up to 20 percentage points. Finally, the comparison with conventional messaging apps confirms that GenAI workloads introduce novel stress factors, such as sustained upstream activity and high-rate bursts, with direct implications for capacity planning and network management. We publicly release the datasets to support reproducibility and foster extensions to other use cases.

Localizing knowledge within Large Language Models (LLMs) is crucial for interpreting their mechanisms and outcomes. Whereas knowledge attribution has so far provided local sample-level explanations, in this work we argue that whenever LLMs are used for classification tasks, a class-level explanation is preferable. We therefore define broader concept areas, i.e., regions of the LLM comprising a small set of neurons that contains the most salient knowledge pertaining to each class and propose methods to identify such areas. We apply our methodology to BERT-based LLMs fine-tuned for downstream classification tasks such as sentiment analysis and attack classification: our results show that it is possible to (i) identify crucial sets of neurons that determine the behaviour of fine-tuned LLMs for explanation purposes, as well as (ii) exploit such concept areas to improve their classification outcomes-yielding up to 6% macro F1-Score improvement on sentiment analysis (public dataset) and 2% on attack classification (private dataset) without requiring further fine-tuning.

In the ever-changing digital environment, ensuring the ongoing effectiveness of traffic analysis and security measures is crucial. Therefore, Class Incremental Learning (CIL) in encrypted Traffic Classification (TC) is essential for adapting to evolving network behaviors and the rapid development of new applications. However, the application of CIL techniques in the TC domain is not straightforward, usually leading to unsatisfactory performance figures. Specifically, the improvement goal is to reduce forgetting on old apps and increase the capacity in learning new ones, in order to improve overall classification performance— reducing the drop from a model “trained-from-scratch”. The contribution of this work is the design of a novel fine-tuning approach called MEMENTO, which is obtained through the careful design of different building blocks: memory management, model training, and rectification strategies. In detail, we propose the application of traffic biflows augmentation strategies to better capitalize on old apps biflows, we introduce improvements in the distillation stage, and we design a general rectification strategy that includes several existing proposals. To assess our proposal, we leverage two publicly-available encrypted network traffic datasets, i.e., MIRAGE19 and CESNET-TLS22. As a result, on both datasets MEMENTO achieves a significant improvement in classifying new apps (w.r.t. the best-performing alternative, i.e., BiC) while maintaining stable performance on old ones. Equally important, MEMENTO achieves satisfactory overall TC performance, filling the gap toward a trained-from-scratch model and offering a considerable gain in terms of time (up to 10× speed-up) to obtain up-to-date and running classifiers. The experimental evaluation relies on a comprehensive performance evaluation workbench for CIL proposals, which is based on a wider set of metrics (as opposed to the existing literature in TC).

Traffic Classification (TC) is experiencing a renewed interest, fostered by the growing popularity of Deep Learning (DL) approaches. In exchange for their proved effectiveness, DL models are characterized by a computationally-intensive training procedure that badly matches the fast-paced release of new (mobile) applications, resulting in significantly limited efficiency of model updates. To address this shortcoming, in this work we systematically explore Class Incremental Learning (CIL) techniques, aimed at adding new apps/services to pre-existing DL-based traffic classifiers without a full retraining, hence speeding up the model’s updates cycle. We investigate a large corpus of state-of-the-art CIL approaches for the DL-based TC task, and delve into their working principles to highlight relevant insight, aiming to understand if there is a case for CIL in TC. We evaluate and discuss their performance varying the number of incremental learning episodes, and the number of new apps added for each episode. Our evaluation is based on the publicly available$\mathtt {MIRAGE19}$dataset comprising traffic of 40 popular Android applications, fostering reproducibility. Despite our analysis reveals their infancy, CIL techniques are a promising research area on the roadmap towards automated DL-based traffic analysis systems.

The promise of Deep Learning (DL) in solving hard problems such as network Traffic Classification (TC) is being held back by the severe lack of transparency and explainability of this kind of approaches. To cope with this strongly felt issue, the field of eXplainable Artificial Intelligence (XAI) has been recently founded, and is providing effective techniques and approaches. Accordingly, in this work we investigate interpretability via XAIbased techniques to understand and improve the behavior of state-of-the-art multimodal and multitask DL traffic classifiers. Using a publicly available security-related dataset (ISCX VPNNONVPN), we explore and exploit XAI techniques to characterize the considered classifiers providing global interpretations (rather than sample-based ones), and define a novel classifier, DISTILLER-EVOLVED, optimized along three objectives: performance, reliability, feasibility. The proposed methodology proves as highly appealing, allowing to much simplify the architecture to get faster training time and shorter classification time, as fewer packets must be collected. This is at the expenses of negligible (or even positive) impact on classification performance, while understanding and controlling the interplay between inputs, model complexity, performance, and reliability.

With the increasing popularity of mobile-app services, malicious software is increasing as well. Accordingly, the interest of the scientific community in Machine and Deep Learning solutions for detecting and classifying malware traffic is growing. In this work, we provide a fair assessment of the performance of a number of data-driven strategies to detect and classify Android malware traffic. Three models are taken into account (Decision Tree, Random Forest, and 1-D Convolutional Neural Network) considering both flat (i.e. non-hierarchical) and hierarchical approaches. The experimental analysis performed using a state-of-art dataset (CIC-AAGM2017) reports that Random Forest exhibits the best performance in a flat setup, while moving to a hierarchical approach could cause significant variation in precision and recall. Such results push for further investigating advanced hierarchical setups and learning schemes.

Traffic Classification (TC), i.e. the collection of procedures for inferring applications and/or services generating network traffic, represents the workhorse for service management and the enabler for valuable profiling information. Sadly, the growing trend toward encrypted protocols (e.g. TLS) and the evolving nature of network traffic make TC design solutions based on payload-inspection and machine learning, respectively, unsuitable. Conversely, Deep Learning (DL) is currently foreseen as a viable means to design traffic classifiers based on automatically-extracted features, reflecting the complex patterns distilled from the multifaceted (encrypted) traffic nature, implicitly carrying information in "multimodal" fashion. To this end, in this paper a novel multimodal DL approach for multitask TC is explored. The latter is able to capitalize traffic data heterogeneity (by learning both intra- and inter-modality dependencies), overcome performance limitations of existing (myopic) single-modality DL-based TC proposals, and solve different traffic categorization problems associated with different providers’ desiderata. Based on a real dataset of encrypted traffic, we report performance gains of our proposal over (a) state-of-art multitask DL architectures and (b) multitask extensions of single-task DL baselines (both based on single-modality philosophy).

The increasing diffusion of mobile devices has dramatically changed the network traffic landscape, with Traffic Classification (TC) surging into a fundamental role while facing new and unprecedented challenges. The recent and appealing adoption of Deep Learning (DL) techniques has risen as the solution overcoming the performance of ML techniques based on tedious and time-consuming handcrafted feature design. Still, the black-box nature of DL models prevents its practical and trustful adoption in critical scenarios where the reliability/interpretation of results/policies is of key importance. To cope with these limitations, eXplainable Artificial Intelligence (XAI) techniques have recently acquired the interest of the community. Accordingly, in this work we investigate trustworthiness and interpretability via XAI-based techniques to understand, interpret and improve the behavior of state-of-the-art multimodal DL traffic classifiers. The proposed methodology, as opposed to common results seen in XAI, attempts to provide global interpretation, rather than sample-based ones. Results, based on an open dataset, allow to complement the above findings with domain knowledge.