TIW 2026 Abstracts

Area 1 - Trustworthy Identity and Interactions on the Web
Full Papers Short Papers

Area 1 - Trustworthy Identity and Interactions on the Web

Full Papers
Paper Nr: 5
Title:

A Method for Android Malware Detection and Possible Malicious Payload Identification through Explainable Transformer Models
Authors:

Francesco Mercaldo, Sebastian Schrittwieser, Patrick Kochberger, Antonella Santone, Paul Tavolato and Fabio Martinelli
Abstract: Malware detection, particularly in mobile operating systems, remains a significant challenge due to the limited effectiveness of both free and commercial signature-based antimalware software in identifying zero-day threats. Signature-based approaches can only detect malware whose signatures have already been incorporated into their databases. Consequently, newly emerging malware must become sufficiently widespread before it can be reliably recognized by existing security solutions. In this paper, we propose a static mal-ware detection method for mobile environments, with a particular focus on the Android platform (the most widespread one). The proposed method exploits four Transformer architectures with the aim to distinguish between malicious and benign Android applications based on their opcode sequences. To overcome the lack of explainability of deep learning models, we integrate two explainability techniques i.e., Integrated Gradients and SHapley Additive exPlanations, to provide token-level explanations and identify the opcodes that contribute most significantly to the malware detection prediction i.e., the malicious payload from the model point of view. Experimental evaluations conducted on a real-world dataset composed by malicious and trusted Android applications demonstrate the effectiveness of the proposed approach. Furthermore, the explainability analysis reveals the opcode patterns that the models associate with malicious behavior, providing valuable insights into the malware detection model decision-making process. These findings suggest that the adoption of Transformer-based architectures with explainability techniques can offer an accurate and trustworthy solution that can support both end users and cybersecurity analysts in Android malware detection.
Download
Paper Nr: 7
Title:

Detecting Adversarial Attacks in Face Recognition through Embedding-Based Reconstruction
Authors:

Sara Lazzaro and Vincenzo De Angelis
Abstract: Face recognition systems support identity verification and identification in settings such as automated border control, access to restricted facilities, and searches in large identity databases. However, these systems can be misled by adversarial perturbations that preserve the visible identity of a person while altering the representation used for automatic matching. It is therefore important to detect suspicious query images before accepting the recognition result. In this paper, we propose a detector that compares each query image with a face reconstructed from the corresponding query embedding. The method measures whether the reconstruction preserves the representation of the query image and its relation to the enrolled gallery, and combines these consistency signals into an adversarial risk score. We evaluate the approach on the LFW dataset using the white-box FGSM and PGD attacks, each considered in both targeted and untargeted variants. At a 5% false-positive rate, the detector achieves 87.25% recall, 98.59% precision, and a 92.57% F1 score. It detects between 93% and 100% of the examples for three of the four evaluated attack settings, while targeted PGD remains more challenging, with a 57% detection rate. These results clearly show that reconstruction consistency provides a useful signal for detecting adversarial query images in face recognition.
Download
Paper Nr: 8
Title:

Verify Once (Possibly Automatically) and Comply Twice: A Case Study at the Cross-Road of the NIS2 and the Italian Cloud Security Regulation
Authors:

Andrea Bisegna, Laura Cristiano, Giovanna Gagliardo, Luca Piras and Silvio Ranise
Abstract: Many Italian Public Administrations are simultaneously bound by the Network and Information Security Directive (NIS2) and by a national Regulation on cloud services issued by the Italian National Cybersecurity Agency (Agenzia per la Cybersicurezza Nazionale, ACN), which risks duplicating compliance effort across separate auditing workflows. The ACN Regulation, like other important EU legislations (e.g., GDPR), presents significant challenges for organisations: its length, complexity, and high-level legal language make it difficult to understand and manage how to comply with its obligations. Furthermore, it prescribes neither a verification methodology nor supporting tools, and its connection with NIS2, which it invokes, is left implicit. We propose a methodology that transforms these obligations into verifiable requirements and, in doing so, makes explicit their correspondence with the NIS2 risk management measures, covering 9 of the 10 measures. Our main finding is that 42 of the 47 test cases we derive are automatable, meaning that tools can support their verification, though for 34 of them the final evaluation against the adopted policies still requires human judgement; only 5 test cases depend entirely on documentary or organisational evidence.
Download
Paper Nr: 9
Title:

Assessing Manipulative Behaviours of AI-Powered Shopping Chatbots
Authors:

Giampaolo Bella, Davide Bonaventura, Andrea Cardilicchia and Viviana Condorelli
Abstract: The increasing adoption of AI-powered shopping chatbots is transforming customer interactions in e-commerce, raising concerns about their potential to influence purchasing decisions in ways that may undermine consumer autonomy. While previous research has primarily investigated manipulative interface design, comparatively little attention has been devoted to conversational manipulation exhibited by AI-powered chatbots. This paper proposes a systematic and reproducible 5-step methodology for assessing manipulative conversational behaviours in AI-powered shopping chatbots. The methodology encompasses chatbot selection, prompt design, prompt instantiation, interaction collection, and behaviour assessment. We apply the methodology to 9 AI-powered shopping chatbots deployed by major e-commerce platforms, collecting a total of 180 interactions. In particular, results show that manipulative conversational behaviours are not uniformly distributed across business objectives. Purchase persuasion and cross-selling exhibit generally limited manipulative tendencies, whereas stronger alignment with manipulative objectives emerges for upselling and, most notably, personal data solicitation.
Download
Paper Nr: 13
Title:

BHASL: A Baseline Hardened Automotive Software Library
Authors:

Giampaolo Bella, Davide Bonaventura, Gianpietro Castiglione, Sergio Esposito, Mirko Giuseppe Mangano, Pierpaolo Pecoraio, Mario Raciti, Salvatore Riccobene, Daniele Francesco Santamaria and Andrea Ventura
Abstract: This work bridges the gap between security and privacy measures as defined by abstract regulatory frameworks, on the one hand, and the technical implementation of such measures in the domain of Connected and Automated Vehicles (CAVs), on the other. We present the software library BHASL, the Baseline Hardened Automotive Software Library, designed to enforce security and privacy controls by design in vehicular systems, in full alignment with the AUTOSAR Adaptive Platform. BHASL operationalises specific controls from ISO/IEC 27002:2022 and the GDPR, integrating Hardware Security Modules (HSM) for cryptographic management alongside architectural support for Data Subject rights, such as the Right to Access and Erasure. Performance evaluation confirms that the computation overhead of BHASL is modest, hence this work provides a verifiable blueprint for systematically validating compliance and protection mechanisms in next-generation automotive software stacks.
Download
Paper Nr: 16
Title:

Towards Confidential Smart Policies Using Fully Homomorphic Encryption
Authors:

Stefano Bistarelli, Chiara Luchini, Ivan Mercanti and Francesco Santini
Abstract: Blockchain-based Smart Policies enable decentralized and auditable policy enforcement through smart contracts. However, the transparency of public blockchains exposes both policy inputs and evaluation outcomes, limiting their applicability to scenarios involving sensitive data. In this paper, we propose a privacy-preserving Smart Policy (SP) framework that integrates Fully Homomorphic Encryption (FHE) within a Self-Sovereign Identity (SSI) ecosystem. Our approach enables policy evaluation directly over encrypted user attributes derived from Verifiable Credentials (VCs), without revealing the underlying information to blockchain participants. The proposed framework leverages FHE in the Ethereum Virtual Machine (EVM), called FHEVM, to execute policy constraints on encrypted data while preserving the transparency and decentralization properties of blockchain systems. We demonstrate the approach’s feasibility through an academic access-control scenario in which students prove eligibility for a service using encrypted attributes such as role, grade average, and enrollment year. Policy constraints are evaluated homomorphically on-chain, and only the contract owner can access the final decision.
Download
Short Papers
Paper Nr: 17
Title:

ClearScreen: Reclaiming end User’s Information Space through Sense-Making and Screening of Unwanted Social Media Content at Endpoints
Authors:

Maria Papadaki
Abstract: Social media users are often subjected to passive content consumption by opaque social media algorithms, often irrespective of their choice. This makes them vulnerable to disinformation and manipulation techniques, especially given the rise of generative AI. Centralised filtering controls can remove harmful content, but can threaten democratic values and human rights, if taken too far. Therefore, there is space for end user controls, which will empower users to reclaim control of their information space and reduce their threat surface. This paper aims to propose ClearScreen, a prototype social media video content screening browser extension based on sense-making, which allows users to automatically screen social media comments and metadata linked to video content for signs of disinformation, harmful or unwanted content. Flagged video content is paused and logged, and can be avoided thereafter, whereas all other content resumes playback. Users can still revoke the decision by choosing to play the video. They can also customise the default screening criteria. ClearScreen is designed to run in the background, focusing on performance to ensure that latency remains within acceptable margins and does not break the video feed. Preliminary results demonstrate that the proposed screening functionality stays within 2000ms latency margins.
Download