Mauro Conti

Stefanos Koffas, Luca Pajola, Stjepan Picek et al.

This work explores stylistic triggers for backdoor attacks in the audio domain: dynamic transformations of malicious samples through guitar effects. We first formalize stylistic triggers - currently missing in the literature. Second, we explore how to develop stylistic triggers in the audio domain by proposing JingleBack. Our experiments confirm the effectiveness of the attack, achieving a 96% attack success rate. Our code is available in https://github.com/skoffas/going-in-style.

Manuel Suarez-Roman, Francesco Marchiori, Mauro Conti et al.

Despite the high volume of open-source Cyber Threat Intelligence (CTI), our understanding of long-term threat actor-victim dynamics remains fragmented due to inconsistent reporting standards and the lack of structured datasets containing comprehensive analytic information. In this paper, we present a large-scale automated analysis of open-source CTI reports spanning two decades. We develop a high-precision, LLM-based pipeline to ingest and structure 16,096 reports, extracting key entities such as attributed threat actors, motivations, victims, reporting vendors, and technical indicators (IoCs and TTPs). Our analysis quantifies the evolution of CTI information density and specialization, characterizing patterns that relate specific threat actors to motivations and victim profiles. Furthermore, we perform a meta-analysis of the CTI industry itself. We identify a fragmented ecosystem of distinct silos where vendors demonstrate significant geographic and sectoral reporting biases. Our marginal coverage analysis reveals that intelligence overlap between vendors is typically low: while a few core providers may offer broad situational awareness, additional sources yield diminishing returns. Overall, our findings characterize the structural biases inherent in the CTI ecosystem, enabling practitioners and researchers to better evaluate the completeness of their intelligence sources.

Giovanni Apruzzese, Luca Pajola, Mauro Conti

Enhancing Network Intrusion Detection Systems (NIDS) with supervised Machine Learning (ML) is tough. ML-NIDS must be trained and evaluated, operations requiring data where benign and malicious samples are clearly labelled. Such labels demand costly expert knowledge, resulting in a lack of real deployments, as well as on papers always relying on the same outdated data. The situation improved recently, as some efforts disclosed their labelled datasets. However, most past works used such datasets just as a 'yet another' testbed, overlooking the added potential provided by such availability. In contrast, we promote using such existing labelled data to cross-evaluate ML-NIDS. Such approach received only limited attention and, due to its complexity, requires a dedicated treatment. We hence propose the first cross-evaluation model. Our model highlights the broader range of realistic use-cases that can be assessed via cross-evaluations, allowing the discovery of still unknown qualities of state-of-the-art ML-NIDS. For instance, their detection surface can be extended--at no additional labelling cost. However, conducting such cross-evaluations is challenging. Hence, we propose the first framework, XeNIDS, for reliable cross-evaluations based on Network Flows. By using XeNIDS on six well-known datasets, we demonstrate the concealed potential, but also the risks, of cross-evaluations of ML-NIDS.

Ehsan Nowroozi, Abhishek, Mohammadreza Mohammadi et al.

Malicious advertisement URLs pose a security risk since they are the source of cyber-attacks, and the need to address this issue is growing in both industry and academia. Generally, the attacker delivers an attack vector to the user by means of an email, an advertisement link or any other means of communication and directs them to a malicious website to steal sensitive information and to defraud them. Existing malicious URL detection techniques are limited and to handle unseen features as well as generalize to test data. In this study, we extract a novel set of lexical and web-scrapped features and employ machine learning technique to set up system for fraudulent advertisement URLs detection. The combination set of six different kinds of features precisely overcome the obfuscation in fraudulent URL classification. Based on different statistical properties, we use twelve different formatted datasets for detection, prediction and classification task. We extend our prediction analysis for mismatched and unlabelled datasets. For this framework, we analyze the performance of four machine learning techniques: Random Forest, Gradient Boost, XGBoost and AdaBoost in the detection part. With our proposed method, we can achieve a false negative rate as low as 0.0037 while maintaining high accuracy of 99.63%. Moreover, we devise a novel unsupervised technique for data clustering using K- Means algorithm for the visual analysis. This paper analyses the vulnerability of decision tree-based models using the limited knowledge attack scenario. We considered the exploratory attack and implemented Zeroth Order Optimization adversarial attack on the detection models.

Serena Tardelli, Leonardo Nizzoli, Maurizio Tesconi et al.

Large-scale online campaigns, malicious or otherwise, require a significant degree of coordination among participants, which sparked interest in the study of coordinated online behavior. State-of-the-art methods for detecting coordinated behavior perform static analyses, disregarding the temporal dynamics of coordination. Here, we carry out the first dynamic analysis of coordinated behavior. To reach our goal we build a multiplex temporal network and we perform dynamic community detection to identify groups of users that exhibited coordinated behaviors in time. Thanks to our novel approach we find that: (i) coordinated communities feature variable degrees of temporal instability; (ii) dynamic analyses are needed to account for such instability, and results of static analyses can be unreliable and scarcely representative of unstable communities; (iii) some users exhibit distinct archetypal behaviors that have important practical implications; (iv) content and network characteristics contribute to explaining why users leave and join coordinated communities. Our results demonstrate the advantages of dynamic analyses and open up new directions of research on the unfolding of online debates, on the strategies of coordinated communities, and on the patterns of online influence.

Mauro Conti, Jiaxin Li, Stjepan Picek et al.

Graph Neural Networks (GNNs), inspired by Convolutional Neural Networks (CNNs), aggregate the message of nodes' neighbors and structure information to acquire expressive representations of nodes for node classification, graph classification, and link prediction. Previous studies have indicated that GNNs are vulnerable to Membership Inference Attacks (MIAs), which infer whether a node is in the training data of GNNs and leak the node's private information, like the patient's disease history. The implementation of previous MIAs takes advantage of the models' probability output, which is infeasible if GNNs only provide the prediction label (label-only) for the input. In this paper, we propose a label-only MIA against GNNs for node classification with the help of GNNs' flexible prediction mechanism, e.g., obtaining the prediction label of one node even when neighbors' information is unavailable. Our attacking method achieves around 60\% accuracy, precision, and Area Under the Curve (AUC) for most datasets and GNN models, some of which are competitive or even better than state-of-the-art probability-based MIAs implemented under our environment and settings. Additionally, we analyze the influence of the sampling method, model selection approach, and overfitting level on the attack performance of our label-only MIA. Both of those factors have an impact on the attack performance. Then, we consider scenarios where assumptions about the adversary's additional dataset (shadow dataset) and extra information about the target model are relaxed. Even in those scenarios, our label-only MIA achieves a better attack performance in most cases. Finally, we explore the effectiveness of possible defenses, including Dropout, Regularization, Normalization, and Jumping knowledge. None of those four defenses prevent our attack completely.

Ying Yuan, Giovanni Apruzzese, Mauro Conti

Existing literature on adversarial Machine Learning (ML) focuses either on showing attacks that break every ML model, or defenses that withstand most attacks. Unfortunately, little consideration is given to the actual feasibility of the attack or the defense. Moreover, adversarial samples are often crafted in the "feature-space", making the corresponding evaluations of questionable value. Simply put, the current situation does not allow to estimate the actual threat posed by adversarial attacks, leading to a lack of secure ML systems. We aim to clarify such confusion in this paper. By considering the application of ML for Phishing Website Detection (PWD), we formalize the "evasion-space" in which an adversarial perturbation can be introduced to fool a ML-PWD -- demonstrating that even perturbations in the "feature-space" are useful. Then, we propose a realistic threat model describing evasion attacks against ML-PWD that are cheap to stage, and hence intrinsically more attractive for real phishers. After that, we perform the first statistically validated assessment of state-of-the-art ML-PWD against 12 evasion attacks. Our evaluation shows (i) the true efficacy of evasion attempts that are more likely to occur; and (ii) the impact of perturbations crafted in different evasion-spaces. Our realistic evasion attempts induce a statistically significant degradation (3-10% at p<0.05), and their cheap cost makes them a subtle threat. Notably, however, some ML-PWD are immune to our most realistic attacks (p=0.22). Finally, as an additional contribution of this journal publication, we are the first to consider the intriguing case wherein an attacker introduces perturbations in multiple evasion-spaces at the same time. These new results show that simultaneously applying perturbations in the problem- and feature-space can cause a drop in the detection rate from 0.95 to 0.

Ehsan Nowroozi, Mohammadreza Mohammadi, Pargol Golmohammadi et al.

In the past decades, the rise of artificial intelligence has given us the capabilities to solve the most challenging problems in our day-to-day lives, such as cancer prediction and autonomous navigation. However, these applications might not be reliable if not secured against adversarial attacks. In addition, recent works demonstrated that some adversarial examples are transferable across different models. Therefore, it is crucial to avoid such transferability via robust models that resist adversarial manipulations. In this paper, we propose a feature randomization-based approach that resists eight adversarial attacks targeting deep learning models in the testing phase. Our novel approach consists of changing the training strategy in the target network classifier and selecting random feature samples. We consider the attacker with a Limited-Knowledge and Semi-Knowledge conditions to undertake the most prevalent types of adversarial attacks. We evaluate the robustness of our approach using the well-known UNSW-NB15 datasets that include realistic and synthetic attacks. Afterward, we demonstrate that our strategy outperforms the existing state-of-the-art approach, such as the Most Powerful Attack, which consists of fine-tuning the network model against specific adversarial attacks. Finally, our experimental results show that our methodology can secure the target network and resists adversarial attack transferability by over 60%.

Pier Paolo Tricomi, Marco Chilese, Mauro Conti et al.

With 1.3 billion users, Instagram (IG) has also become a business tool. IG influencer marketing, expected to generate $33.25 billion in 2022, encourages companies and influencers to create trending content. Various methods have been proposed for predicting a post's popularity, i.e., how much engagement (e.g., Likes) it will generate. However, these methods are limited: first, they focus on forecasting the likes, ignoring the number of comments, which became crucial in 2021. Secondly, studies often use biased or limited data. Third, researchers focused on Deep Learning models to increase predictive performance, which are difficult to interpret. As a result, end-users can only estimate engagement after a post is created, which is inefficient and expensive. A better approach is to generate a post based on what people and IG like, e.g., by following guidelines. In this work, we uncover part of the underlying mechanisms driving IG engagement. To achieve this goal, we rely on statistical analysis and interpretable models rather than Deep Learning (black-box) approaches. We conduct extensive experiments using a worldwide dataset of 10 million posts created by 34K global influencers in nine different categories. With our simple yet powerful algorithms, we can predict engagement up to 94% of F1-Score, making us comparable and even superior to Deep Learning-based method. Furthermore, we propose a novel unsupervised algorithm for finding highly engaging topics on IG. Thanks to our interpretable approaches, we conclude by outlining guidelines for creating successful posts.

Ehsan Nowroozi, Mauro Conti, Yassine Mekdad

In the past decades, the excessive use of the last-generation GAN (Generative Adversarial Networks) models in computer vision has enabled the creation of artificial face images that are visually indistinguishable from genuine ones. These images are particularly used in adversarial settings to create fake social media accounts and other fake online profiles. Such malicious activities can negatively impact the trustworthiness of users identities. On the other hand, the recent development of GAN models may create high-quality face images without evidence of spatial artifacts. Therefore, reassembling uniform color channel correlations is a challenging research problem. To face these challenges, we need to develop efficient tools able to differentiate between fake and authentic face images. In this chapter, we propose a new strategy to differentiate GAN-generated images from authentic images by leveraging spectral band discrepancies, focusing on artificial face image synthesis. In particular, we enable the digital preservation of face images using the Cross-band co-occurrence matrix and spatial co-occurrence matrix. Then, we implement these techniques and feed them to a Convolutional Neural Networks (CNN) architecture to identify the real from artificial faces. Additionally, we show that the performance boost is particularly significant and achieves more than 92% in different post-processing environments. Finally, we provide several research observations demonstrating that this strategy improves a comparable detection method based only on intra-band spatial co-occurrences.

Ehsan Nowroozi, Mohammadreza Mohammadi, Erkay Savas et al.

In the past few years, Convolutional Neural Networks (CNN) have demonstrated promising performance in various real-world cybersecurity applications, such as network and multimedia security. However, the underlying fragility of CNN structures poses major security problems, making them inappropriate for use in security-oriented applications including such computer networks. Protecting these architectures from adversarial attacks necessitates using security-wise architectures that are challenging to attack. In this study, we present a novel architecture based on an ensemble classifier that combines the enhanced security of 1-Class classification (known as 1C) with the high performance of conventional 2-Class classification (known as 2C) in the absence of attacks.Our architecture is referred to as the 1.5-Class (SPRITZ-1.5C) classifier and constructed using a final dense classifier, one 2C classifier (i.e., CNNs), and two parallel 1C classifiers (i.e., auto-encoders). In our experiments, we evaluated the robustness of our proposed architecture by considering eight possible adversarial attacks in various scenarios. We performed these attacks on the 2C and SPRITZ-1.5C architectures separately. The experimental results of our study showed that the Attack Success Rate (ASR) of the I-FGSM attack against a 2C classifier trained with the N-BaIoT dataset is 0.9900. In contrast, the ASR is 0.0000 for the SPRITZ-1.5C classifier.

Marco Alecci, Mauro Conti, Francesco Marchiori et al.

Evasion attacks are a threat to machine learning models, where adversaries attempt to affect classifiers by injecting malicious samples. An alarming side-effect of evasion attacks is their ability to transfer among different models: this property is called transferability. Therefore, an attacker can produce adversarial samples on a custom model (surrogate) to conduct the attack on a victim's organization later. Although literature widely discusses how adversaries can transfer their attacks, their experimental settings are limited and far from reality. For instance, many experiments consider both attacker and defender sharing the same dataset, balance level (i.e., how the ground truth is distributed), and model architecture. In this work, we propose the DUMB attacker model. This framework allows analyzing if evasion attacks fail to transfer when the training conditions of surrogate and victim models differ. DUMB considers the following conditions: Dataset soUrces, Model architecture, and the Balance of the ground truth. We then propose a novel testbed to evaluate many state-of-the-art evasion attacks with DUMB; the testbed consists of three computer vision tasks with two distinct datasets each, four types of balance levels, and three model architectures. Our analysis, which generated 13K tests over 14 distinct attacks, led to numerous novel findings in the scope of transferable attacks with surrogate models. In particular, mismatches between attackers and victims in terms of dataset source, balance levels, and model architecture lead to non-negligible loss of attack performance.

Stefanos Koffas, Stjepan Picek, Mauro Conti

Outsourced training and machine learning as a service have resulted in novel attack vectors like backdoor attacks. Such attacks embed a secret functionality in a neural network activated when the trigger is added to its input. In most works in the literature, the trigger is static, both in terms of location and pattern. The effectiveness of various detection mechanisms depends on this property. It was recently shown that countermeasures in image classification, like Neural Cleanse and ABS, could be bypassed with dynamic triggers that are effective regardless of their pattern and location. Still, such backdoors are demanding as they require a large percentage of poisoned training data. In this work, we are the first to show that dynamic backdoor attacks could happen due to a global average pooling layer without increasing the percentage of the poisoned training data. Nevertheless, our experiments in sound classification, text sentiment analysis, and image classification show this to be very difficult in practice.

Pier Paolo Tricomi, Lisa Facciolo, Giovanni Apruzzese et al.

Did you know that over 70 million of Dota2 players have their in-game data freely accessible? What if such data is used in malicious ways? This paper is the first to investigate such a problem. Motivated by the widespread popularity of video games, we propose the first threat model for Attribute Inference Attacks (AIA) in the Dota2 context. We explain how (and why) attackers can exploit the abundant public data in the Dota2 ecosystem to infer private information about its players. Due to lack of concrete evidence on the efficacy of our AIA, we empirically prove and assess their impact in reality. By conducting an extensive survey on $\sim$500 Dota2 players spanning over 26k matches, we verify whether a correlation exists between a player's Dota2 activity and their real-life. Then, after finding such a link ($p$ < 0.01 and $ρ$ > 0.3), we ethically perform diverse AIA. We leverage the capabilities of machine learning to infer real-life attributes of the respondents of our survey by using their publicly available in-game data. Our results show that, by applyingdomain expertise, some AIA can reach up to 98% precision and over 90% accuracy. This paper hence raises the alarm on a subtle, but concrete threat that can potentially affect the entire competitive gaming landscape. We alerted the developers of Dota2.

Mauro Conti, Jiaxin Li, Stjepan Picek

Membership Inference Attacks (MIAs) infer whether a data point is in the training data of a machine learning model. It is a threat while being in the training data is private information of a data point. MIA correctly infers some data points as members or non-members of the training data. Intuitively, data points that MIA accurately detects are vulnerable. Considering those data points may exist in different target models susceptible to multiple MIAs, the vulnerability of data points under multiple MIAs and target models is worth exploring. This paper defines new metrics that can reflect the actual situation of data points' vulnerability and capture vulnerable data points under multiple MIAs and target models. From the analysis, MIA has an inference tendency to some data points despite a low overall inference performance. Additionally, we implement 54 MIAs, whose average attack accuracy ranges from 0.5 to 0.9, to support our analysis with our scalable and flexible platform, Membership Inference Attacks Platform (VMIAP). Furthermore, previous methods are unsuitable for finding vulnerable data points under multiple MIAs and different target models. Finally, we observe that the vulnerability is not characteristic of the data point but related to the MIA and target model.

Ehsan Nowroozi, Yoosef Habibi, Mauro Conti

The capability of doing effective forensic analysis on printed and scanned (PS) images is essential in many applications. PS documents may be used to conceal the artifacts of images which is due to the synthetic nature of images since these artifacts are typically present in manipulated images and the main artifacts in the synthetic images can be removed after the PS. Due to the appeal of Generative Adversarial Networks (GANs), synthetic face images generated with GANs models are difficult to differentiate from genuine human faces and may be used to create counterfeit identities. Additionally, since GANs models do not account for physiological constraints for generating human faces and their impact on human IRISes, distinguishing genuine from synthetic IRISes in the PS scenario becomes extremely difficult. As a result of the lack of large-scale reference IRIS datasets in the PS scenario, we aim at developing a novel dataset to become a standard for Multimedia Forensics (MFs) investigation which is available at [45]. In this paper, we provide a novel dataset made up of a large number of synthetic and natural printed IRISes taken from VIPPrint Printed and Scanned face images. We extracted irises from face images and it is possible that the model due to eyelid occlusion captured the incomplete irises. To fill the missing pixels of extracted iris, we applied techniques to discover the complex link between the iris images. To highlight the problems involved with the evaluation of the dataset's IRIS images, we conducted a large number of analyses employing Siamese Neural Networks to assess the similarities between genuine and synthetic human IRISes, such as ResNet50, Xception, VGG16, and MobileNet-v2. For instance, using the Xception network, we achieved 56.76\% similarity of IRISes for synthetic images and 92.77% similarity of IRISes for real images.

Ehsan Nowroozi, Yassine Mekdad, Mauro Conti et al.

Recently, the popularity and wide use of the last-generation video conferencing technologies created an exponential growth in its market size. Such technology allows participants in different geographic regions to have a virtual face-to-face meeting. Additionally, it enables users to employ a virtual background to conceal their own environment due to privacy concerns or to reduce distractions, particularly in professional settings. Nevertheless, in scenarios where the users should not hide their actual locations, they may mislead other participants by claiming their virtual background as a real one. Therefore, it is crucial to develop tools and strategies to detect the authenticity of the considered virtual background. In this paper, we present a detection strategy to distinguish between real and virtual video conferencing user backgrounds. We demonstrate that our detector is robust against two attack scenarios. The first scenario considers the case where the detector is unaware about the attacks and inn the second scenario, we make the detector aware of the adversarial attacks, which we refer to Adversarial Multimedia Forensics (i.e, the forensically-edited frames are included in the training set). Given the lack of publicly available dataset of virtual and real backgrounds for video conferencing, we created our own dataset and made them publicly available [1]. Then, we demonstrate the robustness of our detector against different adversarial attacks that the adversary considers. Ultimately, our detector's performance is significant against the CRSPAM1372 [2] features, and post-processing operations such as geometric transformations with different quality factors that the attacker may choose. Moreover, our performance results shows that we can perfectly identify a real from a virtual background with an accuracy of 99.80%.

Marco Arazzi, Mauro Conti, Stefanos Koffas et al.

Federated learning enables collaborative training of machine learning models by keeping the raw data of the involved workers private. Three of its main objectives are to improve the models' privacy, security, and scalability. Vertical Federated Learning (VFL) offers an efficient cross-silo setting where a few parties collaboratively train a model without sharing the same features. In such a scenario, classification labels are commonly considered sensitive information held exclusively by one (active) party, while other (passive) parties use only their local information. Recent works have uncovered important flaws of VFL, leading to possible label inference attacks under the assumption that the attacker has some, even limited, background knowledge on the relation between labels and data. In this work, we are the first (to the best of our knowledge) to investigate label inference attacks on VFL using a zero-background knowledge strategy. To formulate our proposal, we focus on Graph Neural Networks (GNNs) as a target model for the underlying VFL. In particular, we refer to node classification tasks, which are widely studied, and GNNs have shown promising results. Our proposed attack, BlindSage, provides impressive results in the experiments, achieving nearly 100% accuracy in most cases. Even when the attacker has no information about the used architecture or the number of classes, the accuracy remains above 90% in most instances. Finally, we observe that well-known defenses cannot mitigate our attack without affecting the model's performance on the main classification task.

Matteo Gioele Collu, Tom Janssen-Groesbeek, Stefanos Koffas et al.

Large Language Models (LLMs) are being integrated into applications such as chatbots or email assistants. To prevent improper responses, safety mechanisms, such as Reinforcement Learning from Human Feedback (RLHF), are implemented in them. In this work, we bypass these safety measures for ChatGPT, Gemini, and Deepseek by making them impersonate complex personas with personality characteristics that are not aligned with a truthful assistant. First, we create elaborate biographies of these personas, which we then use in a new session with the same chatbots. Our conversations then follow a role-play style to elicit prohibited responses. Using personas, we show that prohibited responses are provided, making it possible to obtain unauthorized, illegal, or harmful information when querying ChatGPT, Gemini, and Deepseek. We show that these chatbots are vulnerable to this attack by getting dangerous information for 40 out of 40 illicit questions in GPT-4.1-mini, Gemini-1.5-flash, 39 out of 40 in GPT-4o-mini, 38 out of 40 in GPT-3.5-turbo, and 2 out of 2 cases in Gemini-2.5-flash and DeepSeek V3. The attack can be carried out manually or automatically using a support LLM, and has proven effective against models deployed between 2023 and 2025.

Sara Bardi, Mauro Conti, Luca Pajola et al.

Social Honeypots are tools deployed in Online Social Networks (OSN) to attract malevolent activities performed by spammers and bots. To this end, their content is designed to be of maximum interest to malicious users. However, by choosing an appropriate content topic, this attractive mechanism could be extended to any OSN users, rather than only luring malicious actors. As a result, honeypots can be used to attract individuals interested in a wide range of topics, from sports and hobbies to more sensitive subjects like political views and conspiracies. With all these individuals gathered in one place, honeypot owners can conduct many analyses, from social to marketing studies. In this work, we introduce a novel concept of social honeypot for attracting OSN users interested in a generic target topic. We propose a framework based on fully-automated content generation strategies and engagement plans to mimic legit Instagram pages. To validate our framework, we created 21 self-managed social honeypots (i.e., pages) on Instagram, covering three topics, four content generation strategies, and three engaging plans. In nine weeks, our honeypots gathered a total of 753 followers, 5387 comments, and 15739 likes. These results demonstrate the validity of our approach, and through statistical analysis, we examine the characteristics of effective social honeypots.

Ehsan Nowroozi, Seyedsadra Seyedshoari, Yassine Mekdad et al.

Digital wallet as a software program or a digital device allows users to conduct various transactions. Hot and cold digital wallets are considered as two types of this wallet. Digital wallets need an online connection fall into the first group, whereas digital wallets can operate without internet connection belong to the second group. Prior to buying a digital wallet, it is important to define for what purpose it will be utilized. The ease with which a mobile phone transaction may be completed in a couple of seconds and the speed with which transactions are executed are reflection of efficiency. One of the most important elements of digital wallets is data organization. Digital wallets are significantly less expensive than classic methods of transaction, which entails various charges and fees. Constantly, demand for their usage is growing due to speed, security, and the ability to conduct transactions between two users without the need of a third party. As the popularity of digital currency wallets grows, the number of security concerns impacting them increases significantly. The current status of digital wallets on the market, as well as the options for an efficient solution for obtaining and utilizing digital wallets. Finally, the digital wallets' security and future improvement prospects are discussed in this chapter.

Gorka Abad, Jing Xu, Stefanos Koffas et al.

Deep learning achieves outstanding results in many machine learning tasks. Nevertheless, it is vulnerable to backdoor attacks that modify the training set to embed a secret functionality in the trained model. The modified training samples have a secret property, i. e., a trigger. At inference time, the secret functionality is activated when the input contains the trigger, while the model functions correctly in other cases. While there are many known backdoor attacks (and defenses), deploying a stealthy attack is still far from trivial. Successfully creating backdoor triggers depends on numerous parameters. Unfortunately, research has not yet determined which parameters contribute most to the attack performance. This paper systematically analyzes the most relevant parameters for the backdoor attacks, i.e., trigger size, position, color, and poisoning rate. Using transfer learning, which is very common in computer vision, we evaluate the attack on state-of-the-art models (ResNet, VGG, AlexNet, and GoogLeNet) and datasets (MNIST, CIFAR10, and TinyImageNet). Our attacks cover the majority of backdoor settings in research, providing concrete directions for future works. Our code is publicly available to facilitate the reproducibility of our results.

Alessandro Lotto, Savio Sciancalepore, Alessandro Brighente et al.

The Manufacturer Usage Description (MUD) standard enables enforcement of network restrictions for IoT devices based on their expected network traffic, as specified by manufacturers in an online MUD file. Devices advertise a URL pointing to this file, yet the standard does not define how to securely bind the issuing device to its profile. As a result, malicious devices can manipulate network policy enforcement by advertising valid URLs referencing genuine MUD profiles, but not intended for that device. Although MUD defines a certificate-based secure issuance method, current deployments rely on the insecure DHCP-based extension due to simpler integration. Existing solutions either depend on Public Key Infrastructure (PKI), break standard compliance, require excessive active manufacturer involvement, or overlook secure profile updates. In this paper, we present FIDEM, a standard-compliant framework for securing DHCP-based MUD URL issuance. FIDEM provides cryptographic binding between IoT devices and their MUD profiles by leveraging Zero-Knowledge-Proof authentication, eliminating PKI reliance, minimizing manufacturers' involvement, and supporting secure profile updates. Formal analysis shows that FIDEM withstands stronger adversaries than in prior work, including supply-chain compromise and attacks using legitimate devices as cryptographic oracles. Our real-world evaluation on two reference constrained devices (ESP32-S3 and ESP32-C6) demonstrates minimal overhead compared to standard DHCP (approximately 5ms and 20mJ) and significant improvements over certificate-based benchmarks (approximately x20 faster, and 35% less energy).

Jiaxin Li, Gorka Abad, Stjepan Picek et al.

Artificial Neural Networks (ANNs), commonly mimicking neurons with non-linear functions to output floating-point numbers, consistently receive the same signals of a data point during its forward time. Unlike ANNs, Spiking Neural Networks (SNNs) get various input signals in the forward time of a data point and simulate neurons in a biologically plausible way, i.e., producing a spike (a binary value) if the accumulated membrane potential of a neuron is larger than a threshold. Even though ANNs have achieved remarkable success in multiple tasks, e.g., face recognition and object detection, SNNs have recently obtained attention due to their low power consumption, fast inference, and event-driven properties. While privacy threats against ANNs are widely explored, much less work has been done on SNNs. For instance, it is well-known that ANNs are vulnerable to the Membership Inference Attack (MIA), but whether the same applies to SNNs is not explored. In this paper, we evaluate the membership privacy of SNNs by considering eight MIAs, seven of which are inspired by MIAs against ANNs. Our evaluation results show that SNNs are more vulnerable (maximum 10% higher in terms of balanced attack accuracy) than ANNs when both are trained with neuromorphic datasets (with time dimension). On the other hand, when training ANNs or SNNs with static datasets (without time dimension), the vulnerability depends on the dataset used. If we convert ANNs trained with static datasets to SNNs, the accuracy of MIAs drops (maximum 11.5% with a reduction of 7.6% on the test accuracy of the target model). Next, we explore the impact factors of MIAs on SNNs by conducting a hyperparameter study. Finally, we show that the basic data augmentation method for static data and two recent data augmentation methods for neuromorphic data can considerably (maximum reduction of 25.7%) decrease MIAs' performance on SNNs.

Filippo Perrina, Francesco Marchiori, Mauro Conti et al.

Cyber Threat Intelligence (CTI) reporting is pivotal in contemporary risk management strategies. As the volume of CTI reports continues to surge, the demand for automated tools to streamline report generation becomes increasingly apparent. While Natural Language Processing techniques have shown potential in handling text data, they often struggle to address the complexity of diverse data sources and their intricate interrelationships. Moreover, established paradigms like STIX have emerged as de facto standards within the CTI community, emphasizing the formal categorization of entities and relations to facilitate consistent data sharing. In this paper, we introduce AGIR (Automatic Generation of Intelligence Reports), a transformative Natural Language Generation tool specifically designed to address the pressing challenges in the realm of CTI reporting. AGIR's primary objective is to empower security analysts by automating the labor-intensive task of generating comprehensive intelligence reports from formal representations of entity graphs. AGIR utilizes a two-stage pipeline by combining the advantages of template-based approaches and the capabilities of Large Language Models such as ChatGPT. We evaluate AGIR's report generation capabilities both quantitatively and qualitatively. The generated reports accurately convey information expressed through formal language, achieving a high recall value (0.99) without introducing hallucination. Furthermore, we compare the fluency and utility of the reports with state-of-the-art approaches, showing how AGIR achieves higher scores in terms of Syntactic Log-Odds Ratio (SLOR) and through questionnaires. By using our tool, we estimate that the report writing time is reduced by more than 40%, therefore streamlining the CTI production of any organization and contributing to the automation of several CTI tasks.

Matteo Gioele Collu, Riccardo Conte, Alberto Giaretta et al.

In this paper, we investigate whether refusal behavior can be predicted from LLM intermediate activations before decoding using linear probes trained on residual stream activations at each transformer block. We find that refusal is linearly decodable well before the final layer, indicating that safety-relevant behavior is represented in intermediate activations before output generation. To test whether this signal is actionable, we introduce Mechanistic AutoDAN, a probe-guided variant of AutoDAN that replaces full-model fitness evaluation with partial forward passes and probe-based scoring inside a genetic prompt search loop. Across the evaluated models, our method achieves attack success rates competitive with vanilla AutoDAN while reducing per-iteration search time by up to 72%, and probe-guided prompts match or exceed AutoDAN's cross-model transfer in several configurations. We further find that the usefulness of probe guidance increases with model scale. Our results show that refusal is not only observable at the output level, but is encoded as a structured and actionable signal in intermediate LLM activations.

Mauro Conti, Nicola Farronato, Stefanos Koffas et al.

Optical Character Recognition (OCR) is a widely used tool to extract text from scanned documents. Today, the state-of-the-art is achieved by exploiting deep neural networks. However, the cost of this performance is paid at the price of system vulnerability. For instance, in backdoor attacks, attackers compromise the training phase by inserting a backdoor in the victim's model that will be activated at testing time by specific patterns while leaving the overall model performance intact. This work proposes a backdoor attack for OCR resulting in the injection of non-readable characters from malicious input images. This simple but effective attack exposes the state-of-the-art OCR weakness, making the extracted text correct to human eyes but simultaneously unusable for the NLP application that uses OCR as a preprocessing step. Experimental results show that the attacked models successfully output non-readable characters for around 90% of the poisoned instances without harming their performance for the remaining instances.

David Polzoni, Tommaso Bianchi, Mauro Conti

Quantum Key Distribution (QKD) is a foundational cryptographic protocol that ensures information-theoretic security. However, classical protocols such as BB84, though favored for their simplicity, offer limited resistance to eavesdropping, and perform poorly under realistic noise conditions. Recent research has explored the use of discrete-time Quantum Walks (QWs) to enhance QKD schemes. In this work, we specifically focus on a one-way QKD protocol, where security depends exclusively on the underlying Quantum Walk (QW) topology, rather than the details of the protocol itself. Our paper introduces a novel protocol based on QWs over a hypercube topology and demonstrates that, under identical parameters, it provides significantly enhanced security and noise resistance compared to the circular topology (i.e., state-of-the-art), thereby strengthening protection against eavesdropping. Furthermore, we introduce an efficient and extensible simulation framework for one-way QKD protocols based on QWs, supporting both circular and hypercube topologies. Implemented with IBM's software development kit for quantum computing (i.e., Qiskit), our toolkit enables noise-aware analysis under realistic noise models. To support reproducibility and future developments, we release our entire simulation framework as open-source. This contribution establishes a foundation for the design of topology-aware QKD protocols that combine enhanced noise tolerance with topologically driven security.

Stefano Cecconello, Mauro Conti, Luca Pajola et al.

The pervasive integration of AI has enabled Offensive AI: the exploitation of AI for malicious ends across the cyber-kill chain. A critical manifestation is the user attribute inference attack, where AI infers sensitive Personally Identifiable Information (PII) from innocuous public data. We explore how music streaming ecosystems, where users routinely release public playlists, can be exploited for Offensive AI. To quantify this threat, we developed musicPIIrate. This novel tool leverages deep learning architectures that utilize both standalone data representations and the structural information embedded in a user's playlist collection. Our design explores set-based approaches (e.g., Deep Sets) and methodologies modeling relationships between playlists (e.g., Graph Neural Networks), which we also combine to leverage both perspectives. Our approach addresses feature extraction from unordered, variable-length set data, enabling accurate PII prediction. Empirical evaluation demonstrates that musicPIIrate achieves state-of-the-art inference accuracy. The tool successfully infers a wide array of attributes, including: Demographics (Age, Country, Gender), Habits (Alcohol, Smoke, Sport), and Personality Traits (OCEAN scores). musicPIIrate outperforms existing methods, beating baselines in 9 out of 15 attribute inference tasks. To counter this vulnerability, we propose JamShield, a lightweight defensive framework. JamShield strategically injects dummy playlists into an account to dilute the PII-carrying signal. Our analysis indicates that JamShield represents a promising defense, lowering inference F1-scores by an average of 10%. This work provides an initial Offensive-AI benchmark for playlist-based PII inference using architectures that leverage set- and graph-structured data and introduces a defense showing encouraging mitigation effects.

Jiaxin Li, Marco Arazzi, Antonino Nocera et al.

Source Inference Attack (SIA) in Federated Learning (FL) aims to identify which client used a target data point for local model training. It allows the central server to audit clients' data usage. In cross-silo FL, a client (silo) collects data from multiple subjects (e.g., individuals, writers, or devices), posing a risk of subject information leakage. Subject Membership Inference Attack (SMIA) targets this scenario and attempts to infer whether any client utilizes data points from a target subject in cross-silo FL. However, existing results on SMIA are limited and based on strong assumptions on the attack scenario. Therefore, we propose a Subject-Level Source Inference Attack (SLSIA) by removing critical constraints that only one client can use a target data point in SIA and imprecise detection of clients utilizing target subject data in SMIA. The attacker, positioned on the server side, controls a target data source and aims to detect all clients using data points from the target subject. Our strategy leverages a binary attack classifier to predict whether the embeddings returned by a local model on test data from the target subject include unique patterns that indicate a client trains the model with data from that subject. To achieve this, the attacker locally pre-trains models using data derived from the target subject and then leverages them to build a training set for the binary attack classifier. Our SLSIA significantly outperforms previous methods on three datasets. Specifically, SLSIA achieves a maximum average accuracy of 0.88 over 50 target subjects. Analyzing embedding distribution and input feature distance shows that datasets with sparse subjects are more susceptible to our attack. Finally, we propose to defend our SLSIA using item-level and subject-level differential privacy mechanisms.

Marco Simoni, Andrea Saracino, Vinod P. et al.

In this paper, we introduce MoRSE (Mixture of RAGs Security Experts), the first specialised AI chatbot for cybersecurity. MoRSE aims to provide comprehensive and complete knowledge about cybersecurity. MoRSE uses two RAG (Retrieval Augmented Generation) systems designed to retrieve and organize information from multidimensional cybersecurity contexts. MoRSE differs from traditional RAGs by using parallel retrievers that work together to retrieve semantically related information in different formats and structures. Unlike traditional Large Language Models (LLMs) that rely on Parametric Knowledge Bases, MoRSE retrieves relevant documents from Non-Parametric Knowledge Bases in response to user queries. Subsequently, MoRSE uses this information to generate accurate answers. In addition, MoRSE benefits from real-time updates to its knowledge bases, enabling continuous knowledge enrichment without retraining. We have evaluated the effectiveness of MoRSE against other state-of-the-art LLMs, evaluating the system on 600 cybersecurity specific questions. The experimental evaluation has shown that the improvement in terms of relevance and correctness of the answer is more than 10\% compared to known solutions such as GPT-4 and Mixtral 7x8.

Matteo Gioele Collu, Umberto Salviati, Roberto Confalonieri et al.

Large Language Models (LLMs) are increasingly being integrated into the scientific peer-review process, raising new questions about their reliability and resilience to manipulation. In this work, we investigate the potential for hidden prompt injection attacks, where authors embed adversarial text within a paper's PDF to influence the LLM-generated review. We begin by formalising three distinct threat models that envision attackers with different motivations -- not all of which implying malicious intent. For each threat model, we design adversarial prompts that remain invisible to human readers yet can steer an LLM's output toward the author's desired outcome. Using a user study with domain scholars, we derive four representative reviewing prompts used to elicit peer reviews from LLMs. We then evaluate the robustness of our adversarial prompts across (i) different reviewing prompts, (ii) different commercial LLM-based systems, and (iii) different peer-reviewed papers. Our results show that adversarial prompts can reliably mislead the LLM, sometimes in ways that adversely affect a "honest-but-lazy" reviewer. Finally, we propose and empirically assess methods to reduce detectability of adversarial prompts under automated content checks.

Denis Donadel, Francesco Marchiori, Luca Pajola et al.

Recent advancements in Artificial Intelligence, and particularly Large Language Models (LLMs), offer promising prospects for aiding system administrators in managing the complexity of modern networks. However, despite this potential, a significant gap exists in the literature regarding the extent to which LLMs can understand computer networks. Without empirical evidence, system administrators might rely on these models without assurance of their efficacy in performing network-related tasks accurately. In this paper, we are the first to conduct an exhaustive study on LLMs' comprehension of computer networks. We formulate several research questions to determine whether LLMs can provide correct answers when supplied with a network topology and questions on it. To assess them, we developed a thorough framework for evaluating LLMs' capabilities in various network-related tasks. We evaluate our framework on multiple computer networks employing proprietary (e.g., GPT4) and open-source (e.g., Llama2) models. Our findings in general purpose LLMs using a zero-shot scenario demonstrate promising results, with the best model achieving an average accuracy of 79.3%. Proprietary LLMs achieve noteworthy results in small and medium networks, while challenges persist in comprehending complex network topologies, particularly for open-source models. Moreover, we provide insight into how prompt engineering can enhance the accuracy of some tasks.

Francesco Marchiori, Mauro Conti

Lithium-ion (Li-ion) batteries are the primary power source in various applications due to their high energy and power density. Their market was estimated to be up to 48 billion U.S. dollars in 2022. However, the widespread adoption of Li-ion batteries has resulted in counterfeit cell production, which can pose safety hazards to users. Counterfeit cells can cause explosions or fires, and their prevalence in the market makes it difficult for users to detect fake cells. Indeed, current battery authentication methods can be susceptible to advanced counterfeiting techniques and are often not adaptable to various cells and systems. In this paper, we improve the state of the art on battery authentication by proposing two novel methodologies, DCAuth and EISthentication, which leverage the internal characteristics of each cell through Machine Learning models. Our methods automatically authenticate lithium-ion battery models and architectures using data from their regular usage without the need for any external device. They are also resilient to the most common and critical counterfeit practices and can scale to several batteries and devices. To evaluate the effectiveness of our proposed methodologies, we analyze time-series data from a total of 20 datasets that we have processed to extract meaningful features for our analysis. Our methods achieve high accuracy in battery authentication for both architectures (up to 0.99) and models (up to 0.96). Moreover, our methods offer comparable identification performances. By using our proposed methodologies, manufacturers can ensure that devices only use legitimate batteries, guaranteeing the operational state of any system and safety measures for the users.

Alessandro Lotto, Francesco Marchiori, Mauro Conti

Modern computing systems inherently trust human input devices, creating an exploitable attack surface for adversarial automation. USB Human Interface Device (HID) emulation attacks, such as those enabled by the USB Rubber Ducky, exploit this assumption to inject arbitrary keystroke sequences while bypassing traditional defenses. Existing countermeasures rely on simple heuristics based on typing speed or timing regularity, which can be easily evaded through basic randomization. Keystroke dynamics analysis offers a more robust alternative by modeling temporal typing behavior. However, prior work frames this problem as behavioral authentication, verifying whether input originates from a specific user rather than detecting automated injection. An alternative approach is continuous monitoring via keylogging integrated with intrusion detection systems, but this requires access to input content, raising significant privacy concerns. In this paper, we provide the first systematic characterization of keystroke dynamics for human-vs-machine discrimination, independent of user identity. Guided by five research questions, we show that robust, privacy-preserving detection is achievable using lightweight models operating solely on timing features, eliminating the need for content access or user profiling. Our analysis reveals that attacker sophistication does not monotonically translate into improved evasion. Instead, robustness depends on exposure to structurally diverse generation strategies rather than increased model complexity. Finally, we quantify the trade-off between detection timeliness and reliability across varying keystroke sequence lengths, identifying practical operating points for early and effective attack interception.

Luca Pajola, Eugenio Caripoti, Stefan Banzer et al.

Every day, our inboxes are flooded with unsolicited emails, ranging between annoying spam to more subtle phishing scams. Unfortunately, despite abundant prior efforts proposing solutions achieving near-perfect accuracy, the reality is that countering malicious emails still remains an unsolved dilemma. This "open problem" paper carries out a critical assessment of scientific works in the context of phishing email detection. First, we focus on the benchmark datasets that have been used to assess the methods proposed in research. We find that most prior work relied on datasets containing emails that -- we argue -- are not representative of current trends, and mostly encompass the English language. Based on this finding, we then re-implement and re-assess a variety of detection methods reliant on machine learning (ML), including large-language models (LLM), and release all of our codebase -- an (unfortunately) uncommon practice in related research. We show that most such methods achieve near-perfect performance when trained and tested on the same dataset -- a result which intrinsically hinders development (how can future research outperform methods that are already near perfect?). To foster the creation of "more challenging benchmarks" that reflect current phishing trends, we propose E-PhishGEN, an LLM-based (and privacy-savvy) framework to generate novel phishing-email datasets. We use our E-PhishGEN to create E-PhishLLM, a novel phishing-email detection dataset containing 16616 emails in three languages. We use E-PhishLLM to test the detectors we considered, showing a much lower performance than that achieved on existing benchmarks -- indicating a larger room for improvement. We also validate the quality of E-PhishLLM with a user study (n=30). To sum up, we show that phishing email detection is still an open problem -- and provide the means to tackle such a problem by future research.

Denis Donadel, Gabriele Crestanello, Giulio Morandini et al.

Industrial Control Systems (ICS) manage critical infrastructures like power grids and water treatment plants. Cyberattacks on ICSs can disrupt operations, causing severe economic, environmental, and safety issues. For example, undetected pollution in a water plant can put the lives of thousands at stake. ICS researchers have increasingly turned to honeypots -- decoy systems designed to attract attackers, study their behaviors, and eventually improve defensive mechanisms. However, existing ICS honeypots struggle to replicate the ICS physical process, making them susceptible to detection. Accurately simulating the noise in ICS physical processes is challenging because different factors produce it, including sensor imperfections and external interferences. In this paper, we propose SimProcess, a novel framework to rank the fidelity of ICS simulations by evaluating how closely they resemble real-world and noisy physical processes. It measures the simulation distance from a target system by estimating the noise distribution with machine learning models like Random Forest. Unlike existing solutions that require detailed mathematical models or are limited to simple systems, SimProcess operates with only a timeseries of measurements from the real system, making it applicable to a broader range of complex dynamic systems. We demonstrate the framework's effectiveness through a case study using real-world power grid data from the EPIC testbed. We compare the performance of various simulation methods, including static and generative noise techniques. Our model correctly classifies real samples with a recall of up to 1.0. It also identifies Gaussian and Gaussian Mixture as the best distribution to simulate our power systems, together with a generative solution provided by an autoencoder, thereby helping developers to improve honeypot fidelity. Additionally, we make our code publicly available.

Luca Attanasio, Mauro Conti, Denis Donadel et al.

The impact of global warming and the imperative to limit climate change have stimulated the need to develop new solutions based on renewable energy sources. One of the emerging trends in this endeavor are the Electric Vehicles (EVs), which use electricity instead of traditional fossil fuels as a power source, relying on the Vehicle-to-Grid (V2G) paradigm. The novelty of such a paradigm requires careful analysis to avoid malicious attempts. An attacker can exploit several surfaces, such as the remote connection between the Distribution Grid and Charging Supply or the authentication system between the charging Supply Equipment and the Electric Vehicles. However, V2G architecture's high cost and complexity in implementation can restrain this field's research capability. In this paper, we approach this limitation by proposing MiniV2G, an open-source emulator to simulate Electric Vehicle Charging (EVC) built on top of Mininet and RiseV2G. MiniV2G is particularly suitable for security researchers to study and test real V2G charging scenarios. MiniV2G can reproduce with high fidelity a V2G architecture to easily simulate an EV charging process. Finally, we present a MiniV2G application and show how MiniV2G can be used to study V2G communication and develop attacks and countermeasures that can be applied to real systems. Since we believe our tool can be of great help for research in this field, we also made it freely available.

Parvez Faruki, Hossein Fereidooni, Vijay Laxmi et al.

Mobile devices have become ubiquitous due to centralization of private user information, contacts, messages and multiple sensors. Google Android, an open-source mobile Operating System (OS), is currently the market leader. Android popularity has motivated the malware authors to employ set of cyber attacks leveraging code obfuscation techniques. Obfuscation is an action that modifies an application (app) code, preserving the original semantics and functionality to evade anti-malware. Code obfuscation is a contentious issue. Theoretical code analysis techniques indicate that, attaining a verifiable and secure obfuscation is impossible. However, obfuscation tools and techniques are popular both among malware developers (to evade anti-malware) and commercial software developers (protect intellectual rights). We conducted a survey to uncover answers to concrete and relevant questions concerning Android code obfuscation and protection techniques. The purpose of this paper is to review code obfuscation and code protection practices, and evaluate efficacy of existing code de-obfuscation tools. In particular, we discuss Android code obfuscation methods, custom app protection techniques, and various de-obfuscation methods. Furthermore, we review and analyse the obfuscation techniques used by malware authors to evade analysis efforts. We believe that, there is a need to investigate efficiency of the defense techniques used for code protection. This survey would be beneficial to the researchers and practitioners, to understand obfuscation and de-obfuscation techniques to propose novel solutions on Android.

Sameera K. M., Serena Nicolazzo, Marco Arazzi et al.

Federated Learning (FL) has recently arisen as a revolutionary approach to collaborative training Machine Learning models. According to this novel framework, multiple participants train a global model collaboratively, coordinating with a central aggregator without sharing their local data. As FL gains popularity in diverse domains, security, and privacy concerns arise due to the distributed nature of this solution. Therefore, integrating this strategy with Blockchain technology has been consolidated as a preferred choice to ensure the privacy and security of participants. This paper explores the research efforts carried out by the scientific community to define privacy solutions in scenarios adopting Blockchain-Enabled FL. It comprehensively summarizes the background related to FL and Blockchain, evaluates existing architectures for their integration, and the primary attacks and possible countermeasures to guarantee privacy in this setting. Finally, it reviews the main application scenarios where Blockchain-Enabled FL approaches have been proficiently applied. This survey can help academia and industry practitioners understand which theories and techniques exist to improve the performance of FL through Blockchain to preserve privacy and which are the main challenges and future directions in this novel and still under-explored context. We believe this work provides a novel contribution respect to the previous surveys and is a valuable tool to explore the current landscape, understand perspectives, and pave the way for advancements or improvements in this amalgamation of Blockchain and Federated Learning.

Daniele Mari, Saverio Cavasin, Simone Milani et al.

The success of learning-based coding techniques and the development of learning-based image coding standards, such as JPEG-AI, point towards the adoption of such solutions in different fields, including the storage of biometric data, like fingerprints. However, the peculiar nature of learning-based compression artifacts poses several issues concerning their impact and effectiveness on extracting biometric features and landmarks, e.g., minutiae. This problem is utterly stressed by the fact that most models are trained on natural color images, whose characteristics are very different from usual biometric images, e.g, fingerprint or iris pictures. As a matter of fact, these issues are deemed to be accurately questioned and investigated, being such analysis still largely unexplored. This study represents the first investigation about the adaptability of learning-based image codecs in the storage of fingerprint images by measuring its impact on the extraction and characterization of minutiae. Experimental results show that at a fixed rate point, learned solutions considerably outperform previous fingerprint coding standards, like JPEG2000, both in terms of distortion and minutiae preservation. Indeed, experimental results prove that the peculiarities of learned compression artifacts do not prevent automatic fingerprint identification (since minutiae types and locations are not significantly altered), nor do compromise image quality for human visual inspection (as they gain in terms of BD rate and PSNR of 47.8% and +3.97dB respectively).

Ehsan Nowroozi, Imran Haider, Rahim Taheri et al.

Federated Learning (FL) is a machine learning (ML) approach that enables multiple decentralized devices or edge servers to collaboratively train a shared model without exchanging raw data. During the training and sharing of model updates between clients and servers, data and models are susceptible to different data-poisoning attacks. In this study, our motivation is to explore the severity of data poisoning attacks in the computer network domain because they are easy to implement but difficult to detect. We considered two types of data-poisoning attacks, label flipping (LF) and feature poisoning (FP), and applied them with a novel approach. In LF, we randomly flipped the labels of benign data and trained the model on the manipulated data. For FP, we randomly manipulated the highly contributing features determined using the Random Forest algorithm. The datasets used in this experiment were CIC and UNSW related to computer networks. We generated adversarial samples using the two attacks mentioned above, which were applied to a small percentage of datasets. Subsequently, we trained and tested the accuracy of the model on adversarial datasets. We recorded the results for both benign and manipulated datasets and observed significant differences between the accuracy of the models on different datasets. From the experimental results, it is evident that the LF attack failed, whereas the FP attack showed effective results, which proved its significance in fooling a server. With a 1% LF attack on the CIC, the accuracy was approximately 0.0428 and the ASR was 0.9564; hence, the attack is easily detectable, while with a 1% FP attack, the accuracy and ASR were both approximately 0.9600, hence, FP attacks are difficult to detect. We repeated the experiment with different poisoning percentages.

Ying Yuan, Cristiano Alex Rado, Giovanni Apruzzese et al.

Phishing webpages are continuously polluting the Web. Plenty of countermeasures have been proposed and the most advanced techniques leverage machine-learning methods that infer whether a webpage is benign or not by inspecting its visual representation. Yet, despite the demonstrated effectiveness of such detection methods, this class of defenses is, by design, susceptible to a kind of subtle-but-cheap timing-based attacks which -- worryingly, and perhaps surprisingly -- have never been investigated so far. Such an oversight questions the overall reliability of these defenses in the wild. First, we show that timing-based evasion attacks have not been accounted for by prior work on visual phishing websites detectors. Then, we elucidate the intrinsic vulnerability of these detectors: they can be bypassed by delaying the rendering of webpage elements. Practically, these detectors must compute the visual similarity between a target webpage and a known legitimate one. This requires taking a "snapshot" of the target webpage before the similarity computation. Attackers can deliberately delay the rendering of key elements, such as the logo, so that these elements appear fully only after the snapshot has been taken. This simple tactic misleads the visual-similarity module, leading the system to incorrectly classify the phishing page as benign. We empirically show that state-of-the-art detectors can be completely defeated (detection rate dropping from 100% to 0%) by employing easy-to-apply problem-space techniques such as curtain effects. We also carry out a user study, evaluating the effectiveness of these attacks against real humans, and find that end users are unable to reliably identify our "perturbations" (p<.05). Finally, we propose mitigations, including a browser-extension that, without making any call to remote services, warns users that they may have landed on a phishing webpage.

Florine W. Dekker, Zekeriya Erkin, Mauro Conti

Decentralised learning has recently gained traction as an alternative to federated learning in which both data and coordination are distributed. To preserve the confidentiality of users' data, decentralised learning relies on differential privacy, multi-party computation, or both. However, running multiple privacy-preserving summations in sequence may allow adversaries to perform reconstruction attacks. Current reconstruction countermeasures either cannot trivially be adapted to the distributed setting, or add excessive amounts of noise. In this work, we first show that passive honest-but-curious adversaries can infer other users' private data after several privacy-preserving summations. For example, in subgraphs with 18 users, we show that only three passive honest-but-curious adversaries succeed at reconstructing private data 11.0% of the time, requiring an average of 8.8 summations per adversary. The success rate depends only on the adversaries' direct neighbourhood, and is independent of the size of the full network. We consider weak adversaries that do not control the graph topology, cannot exploit the summation's inner workings, and do not have auxiliary knowledge; and show that these adversaries can still infer private data. We analyse how reconstruction relates to topology and propose the first topology-based decentralised defence against reconstruction attacks. We show that reconstruction requires a number of adversaries linear in the length of the network's shortest cycle. Consequently, exact attacks over privacy-preserving summations are impossible in acyclic networks. Our work is a stepping stone for a formal theory of topology-based decentralised reconstruction defences. Such a theory would generalise our countermeasure beyond summation, define confidentiality in terms of entropy, and describe the interactions with (topology-aware) differential privacy.

Emad Efatinasab, Alessandro Brighente, Denis Donadel et al.

Smart grids are crucial for meeting rising energy demands driven by global population growth and urbanization. By integrating renewable energy sources, they enhance efficiency, reliability, and sustainability. However, ensuring their availability and security requires advanced operational control and safety measures. Although artificial intelligence and machine learning can help assess grid stability, challenges such as data scarcity and cybersecurity threats, particularly adversarial attacks, remain. Data scarcity is a major issue, as obtaining real-world instances of grid instability requires significant expertise, resources, and time. Yet, these instances are critical for testing new research advancements and security mitigations. This paper introduces a novel framework for detecting instability in smart grids using only stable data. It employs a Generative Adversarial Network (GAN) where the generator is designed not to produce near-realistic data but instead to generate Out-Of-Distribution (OOD) samples with respect to the stable class. These OOD samples represent unstable behavior, anomalies, or disturbances that deviate from the stable data distribution. By training exclusively on stable data and exposing the discriminator to OOD samples, our framework learns a robust decision boundary to distinguish stable conditions from any unstable behavior, without requiring unstable data during training. Furthermore, we incorporate an adversarial training layer to enhance resilience against attacks. Evaluated on a real-world dataset, our solution achieves up to 98.1\% accuracy in predicting grid stability and 98.9\% in detecting adversarial attacks. Implemented on a single-board computer, it enables real-time decision-making with an average response time of under 7ms.

Matteo Gioele Collu, Tom Janssen-Groesbeek, Stefanos Koffas et al.

Large Language Models (LLMs) are being integrated into applications such as chatbots or email assistants. To prevent improper responses, safety mechanisms, such as Reinforcement Learning from Human Feedback (RLHF), are implemented in them. In this work, we bypass these safety measures for ChatGPT, Gemini, and Deepseek by making them impersonate complex personas with personality characteristics that are not aligned with a truthful assistant. First, we create elaborate biographies of these personas, which we then use in a new session with the same chatbots. Our conversations then follow a role-play style to elicit prohibited responses. Using personas, we show that prohibited responses are provided, making it possible to obtain unauthorized, illegal, or harmful information when querying ChatGPT, Gemini, and Deepseek. We show that these chatbots are vulnerable to this attack by getting dangerous information for 40 out of 40 illicit questions in GPT-4.1-mini, Gemini-1.5-flash, 39 out of 40 in GPT-4o-mini, 38 out of 40 in GPT-3.5-turbo, and 2 out of 2 cases in Gemini-2.5-flash and DeepSeek V3. The attack can be carried out manually or automatically using a support LLM, and has proven effective against models deployed between 2023 and 2025.

Alberto Castagnaro, Umberto Salviati, Mauro Conti et al.

Large Language Models (LLMs) have transformed human-machine interaction since ChatGPT's 2022 debut, with Retrieval-Augmented Generation (RAG) emerging as a key framework that enhances LLM outputs by integrating external knowledge. However, RAG's reliance on ingesting external documents introduces new vulnerabilities. This paper exposes a critical security gap at the data loading stage, where malicious actors can stealthily corrupt RAG pipelines by exploiting document ingestion. We propose a taxonomy of 9 knowledge-based poisoning attacks and introduce two novel threat vectors -- Content Obfuscation and Content Injection -- targeting common formats (DOCX, HTML, PDF). Using an automated toolkit implementing 19 stealthy injection techniques, we test five popular data loaders, finding a 74.4% attack success rate across 357 scenarios. We further validate these threats on six end-to-end RAG systems -- including white-box pipelines and black-box services like NotebookLM and OpenAI Assistants -- demonstrating high success rates and critical vulnerabilities that bypass filters and silently compromise output integrity. Our results emphasize the urgent need to secure the document ingestion process in RAG systems against covert content manipulations.

Francesco Marchiori, Denis Donadel, Alessandro Brighente et al.

Electric Vehicles (EVs) are rapidly gaining adoption as a sustainable alternative to fuel-powered vehicles, making secure charging infrastructure essential. Despite traditional authentication protocols, recent results showed that attackers may steal energy through tailored relay attacks. One countermeasure is leveraging the EV's fingerprint on the current exchanged during charging. However, existing methods focus on the final charging stage, allowing malicious actors to consume substantial energy before being detected and repudiated. This underscores the need for earlier and more effective authentication methods to prevent unauthorized charging. Meanwhile, profiling raises privacy concerns, as uniquely identifying EVs through charging patterns could enable user tracking. In this paper, we propose a framework for uniquely identifying EVs using physical measurements from the early charging stages. We hypothesize that voltage behavior early in the process exhibits similar characteristics to current behavior in later stages. By extracting features from early voltage measurements, we demonstrate the feasibility of EV profiling. Our approach improves existing methods by enabling faster and more reliable vehicle identification. We test our solution on a dataset of 7408 usable charges from 49 EVs, achieving up to 0.86 accuracy. Feature importance analysis shows that near-optimal performance is possible with just 10 key features, improving efficiency alongside our lightweight models. This research lays the foundation for a novel authentication factor while exposing potential privacy risks from unauthorized access to charging data.

Sameera K. M., Vinod P., Anderson Rocha et al.

In the era of data expansion, ensuring data privacy has become increasingly critical, posing significant challenges to traditional AI-based applications. In addition, the increasing adoption of IoT devices has introduced significant cybersecurity challenges, making traditional Network Intrusion Detection Systems (NIDS) less effective against evolving threats, and privacy concerns and regulatory restrictions limit their deployment. Federated Learning (FL) has emerged as a promising solution, allowing decentralized model training while maintaining data privacy to solve these issues. However, despite implementing privacy-preserving technologies, FL systems remain vulnerable to adversarial attacks. Furthermore, data distribution among clients is not heterogeneous in the FL scenario. We propose WeiDetect, a two-phase, server-side defense mechanism for FL-based NIDS that detects malicious participants to address these challenges. In the first phase, local models are evaluated using a validation dataset to generate validation scores. These scores are then analyzed using a Weibull distribution, identifying and removing malicious models. We conducted experiments to evaluate the effectiveness of our approach in diverse attack settings. Our evaluation included two popular datasets, CIC-Darknet2020 and CSE-CIC-IDS2018, tested under non-IID data distributions. Our findings highlight that WeiDetect outperforms state-of-the-art defense approaches, improving higher target class recall up to 70% and enhancing the global model's F1 score by 1% to 14%.

Saverio Cavasin, Pietro Biasetton, Mattia Tamiazzo et al.

During criminal investigations, images of persons of interest directly influence the success of identification procedures. However, law enforcement agencies often face challenges related to the scarcity of high-quality images or their obsolescence, which can affect the accuracy and success of people searching processes. This paper introduces a novel forensic mugshot augmentation framework aimed at addressing these limitations. Our approach enhances the identification probability of individuals by generating additional, high-quality images through customizable data augmentation techniques, while maintaining the biometric integrity and consistency of the original data. Several experimental results show that our method significantly improves identification accuracy and robustness across various forensic scenarios, demonstrating its effectiveness as a trustworthy tool law enforcement applications. Index Terms: Digital Forensics, Person re-identification, Feature extraction, Data augmentation, Visual-Language models.