Chain-of-Thought (CoT) guides large language models to reason step-by-step, yielding remarkable performance gains across diverse tasks. However, this structured reasoning process also introduces novel and underexplored security risks. In this paper, we present an in-depth analysis of fine-tuning attacks targeting CoT-enabled LLMs, with particular focus on “aha moments” during reasoning, which are critical intermediate steps the model takes to make a significant decision or change its behavior. Through experiments on six CoT models and three non-CoT baselines, we find that even aligned CoT models can be more harmful than their base models. Moreover, the reasoning process frequently contains more harmful and actionable content than the final answer, even when the final answer refuses a harmful request. By examining the causal relationship between the reasoning process and the final outputs, we identify two distinct failure modes, Unintentional Leakage and Harmful Escalation, that systematically drive the generation of harmful reasoning. To rigorously assess these risks, we propose an evaluation framework grounded in the EU AI Act and construct a policy-aligned benchmark dataset for CoT reasoning. Our findings expose inherent vulnerabilities in CoT and offer insights for supervising and aligning the reasoning process in LLMs.

Uncrewed Aerial Vehicles (UAVs) are emerging as an integral part of delivering packages, food, and medicines for fast and efficient services. They rely on 5G networks offering high-speed, low-latency, and reliable connectivity for the exchange of mission-critical data. The 5G-connected drones remain vulnerable to cyber security attacks, including those impacting confidentiality, authentication, and integrity. In this paper, we present an edge-assisted data aggregation framework that reduces the drone's computation overhead and allow for secure data sharing between the drones and the Ground Control Station (GCS). The framework relies on a multi-level Attribute-Based Encryption (ABE) enhanced with an aggregate scheme using the Homomorphic Encryption (HE) properties. By integrating the property of HE with ABE, we ensure that only authorised entities can decrypt the encrypted messages under the threshold policy. To reduce the computation overhead at drones, we offload most computationally expensive operations in the encryption and decryption phases to an edge server. Our security analysis demonstrated that the proposed scheme guarantees confidentiality, access control, and key management security while resisting UAV-specific attacks such as eavesdropping, man-in-the-middle attacks and data injection. We validate the proposed scheme using a realistic testbed that includes a Holybro Pixhawk drone and Raspberry Pi. The experimental results demonstrate that the edge-assisted ERAHE framework effectively reduces cryptographic latency and computation burden on UAVs by partitioning expensive operations between the drone and the edge node. ERAHE achieves an optimal balance between cryptographic robustness and lightweight performance, making it well-suited for mission-critical applications.

The convergence of autonomous aerial systems and networking technologies has given rise to the Internet of Drones (IoD) as a compelling paradigm, gaining significant attention from academia and industry stakeholders. Drones often operate in swarm formations to collaboratively achieve autonomous coordination and aerial intelligence, thereby forming a Flying Ad Hoc Network (FANET). However, the persistent vulnerability remains in the insecure communication link, exposing the network to eavesdropping and unauthorized access. Addressing such shortcomings necessitates a robust Authentication and Key Exchange (AKE) protocol. Therefore, we have designed a quantum secure AKE protocol integrating NIST-proven quantum secure primitives, including ML-DSA, symmetric AES, and hash functions. To the best of our knowledge, this is the first AKE protocol that leverages a quantum secure signature scheme for securing IoD-enabled FANET applications. The designed protocol incorporates hardware-specific fingerprinting integrated with a noise tolerance mechanism to eliminate the risk of unauthorized device tampering. The use of re-synchronization and robust security measures for credential management further enhances its resilience against desynchronization and stolen attacks. The findings of performance evaluation exhibit the superiority of the designed protocol over the prevalent AKE protocols, with a remarkable reduction of 67.62% in computation cost while achieving a 50% improvement in overall security. Finally, implementing a complete authentication cycle using PIX32 and Pixhawk 6C drones sets a new benchmark as a practical validation of the designed AKE protocol within a real-world IoD testbed.

The authentication and key establishment (AKE) mechanism is considered one of the promising solutions for securing communication in Internet of Drones (IoD) applications. Nevertheless, existing AKE mechanisms based on traditional cryptographic techniques rely on integer factorization and discrete logarithms, which are no longer safe with the advent of quantum computers. These shortcomings motivate us to design a cutting-edge Quantum Secure Authentication and Key-Establishment mechanism (QSAKE) for the IoD environment. To the best of our knowledge, QSAKE is the pioneered work that uses advanced quantum-safe cryptography, providing a strong defence beyond traditional methods. To further enhance security, it eliminates storing long-term secrets directly in drone memory, reducing the risk of unauthorized access. A Holybro Pixhawk-based microcontroller is used with a Raspberry Pi connected to a Xilinx Arty A7-100T FPGA board to develop a realistic testbed. Finally, this work stands out as a groundbreaking application of a complete authentication process within a practical IoD testbed, demonstrating its high efficacy and practicality.

Autonomous Vehicles (AVs) rely heavily on machine learning (ML) algorithms for real-time perception and decisionmaking, making them vulnerable to adversarial attacks. While prior studies have explored physical attacks and adversarial manipulations on static datasets, there remains a significant gap in understanding how ML-based attacks affect AV performance in dynamic environments. To address this, we simulate two representative adversarial attacks, Fast Gradient Sign Method (FGSM) and Simple Black-box Attack (SimBA), within the CARLA simulator, targeting camera sensors under urban and motorway driving scenarios. Our experiments demonstrate that both attacks can significantly increase the probability of collisions and reduce time-to-collision (TTC), with FGSM and SimBA causing over $60 \%$ collision rates in urban settings and over 95% in motorway scenarios, compared to only 12% and 3% under baseline conditions. Furthermore, our real-time feasibility analysis shows that even the most efficient attacks require at least a $60 \times$ speedup to meet the sub- 50 ms end-to-end latency requirements of AV systems - a threshold dictated by real-time perception needs, where camera frames are typically processed at 30-60 FPS and decisions must be made within tens of milliseconds to ensure safety. These findings underscore the limitations of traditional white-box and black-box attacks in real-time deployments and highlight the need for simulationaware optimizations and hybrid attack strategies. This study provides the first comprehensive evaluation of adversarial attacks on camera sensors in dynamic environments, offering practical insights for both attack refinement and defense development in AV systems.

Unmanned Aerial Vehicles (UAVs), commonly known as drones, have become essential for transporting packages, food, medicines, and other goods due to the growing demand for fast and efficient delivery services. The implementation of 5G technology provides high-speed, low-latency, and reliable connectivity, which allows drones to exchange mission-critical data effectively. However, drones utilizing 5G networks are susceptible to security threats that could compromise essential security requirements such as confidentiality, authentication, integrity, and availability. In this paper, we propose a robust communication framework designed for secure interactions among 5G-connected delivery drones. Our framework relies on a novel Attribute-Based Encryption with Aggregation that is composed by an enhanced multi-level Attribute-Based Encryption (ABE) scheme with Homomorphic Encryption (HE). By integrating HE with the ABE scheme, the Ground Control Station (GCS) and the parent drone can decrypt mission-critical messages as required. This ensures that only authorized entities have access to sensitive data. Additionally, in scenarios that require data aggregation without exposing the underlying content, the HE property within the ABE scheme facilitates this process. As a result, encrypted data subsets can be aggregated anywhere in the network without the need for decryption, thereby preserving data confidentiality and enhancing both communication and computational efficiency. We utilize a hierarchical Chain-Based Data Aggregation (CBDA) model for the structural organization of drones, which enhances communication efficiency and reduces energy consumption. By integrating multi-level ABE for flexible and secure access control with HE, our framework effectively addresses major security challenges faced by 5G-based drone networks, ensuring the security and efficient management of mission-critical data.

With the rapid development of V2X and autonomous driving technologies, the security and stability of platoons based on Connected Autonomous Vehicles (CAVs) have become a critical research focus. However, these systems remain vulnerable when facing sophisticated attacks. In this study, we propose a Game Theory-based attack framework that introduces a Ghost Vehicle capable of executing three distinct attack modes, Leader Attack, Mid-Platoon Attack, and Follower Attack. Experimental results demonstrate that our attack effectively destabilizes the platoon during Leader Attack and Mid-Platoon Attack, while seamlessly integrating as a normal member during the Follower Attack. Unlike prior sensor-spoofing attacks that rely on hardware-intensive methods, the behavior of the Ghost Vehicle is dynamically controlled using Mode Predictive Control (MPC) and Finite State Machine (FSM), ensuring high efficiency with a maximum computation time of 40 seconds, which represents an average across multiple runs. Our attack modes are validated in a highway scenario, demonstrating the high effectiveness with low computational cost.

With the growing presence of drones in our skies, securing their operations and ensuring reliable communication has become more crucial than ever. These drones form interconnected networks known as the Internet of Drones (IoD) to facilitate real-time coordination, autonomous aerial surveillance, and special aerial tasks. However, the interaction between drones and ground stations occurs over unregulated and dynamic communication channels, introducing security vulnerabilities such as impersonation, Man-in-the-Middle (MitM), and forgery attacks. Implementing robust authentication protocols can serve as a promising solution to protect drone operations and communication, thereby ensuring the safety and security of our skies. In this article, we introduce a novel authentication and key establishment protocol that uniquely integrates level-triggered Physically Unclonable Function (PUF), BCH error-correcting code, and AES-GCM symmetric encryption, setting a new standard for secure and reliable communication between drones and ground stations. We demonstrate the robustness of our protocol through comprehensive security verification using the Scyther tool, coupled with formal validation within the Random Oracle Model (ROM). Through rigorous performance analysis, we demonstrate the superiority of our protocol over baseline protocols, achieving 64.37% greater efficiency in computation and 26.03% reduction in communication overheads. We also present a realistic implementation of our protocol using Pix32 v6 companion with Raspberry Pi as drone and laptop device as ground station server. The PUF has been implemented in Xilinx Arty A7-100T FPGA board. To the best of our knowledge, this is the first work demonstrating the implementation of a complete authentication cycle in real-world IoD settings.

The use of voice-controlled virtual assistants (VAs) is significant, and user numbers increase every year. Extensive use of VAs has provided the large, cash-rich technology companies who sell them with another way of consuming users’ data, providing a lucrative revenue stream. Whilst these companies are legally obliged to treat users’ information “fairly and responsibly,” artificial intelligence techniques used to process data have become incredibly sophisticated, leading to users’ concerns that a lack of clarity is making it hard to understand the nature and scope of data collection and use. There has been little work undertaken on a self-contained user awareness tool targeting VAs. PrivExtractor, a novel web-based awareness dashboard for VA users, intends to redress this imbalance of understanding between the data “processors” and the user. It aims to achieve this using the four largest VA vendors as a case study and providing a comparison function that examines the four companies’ privacy practices and their compliance with data protection law. As a result of this research, we conclude that the companies studied are largely compliant with the law, as expected. However, the user remains disadvantaged due to the ineffectiveness of current data regulation that does not oblige the companies to fully and transparently disclose how and when they use, share, or profit from the data. Furthermore, the software tool developed during the research is, we believe, the first that is capable of a comparative analysis of VA privacy with a visual demonstration to increase ease of understanding for the user.

The Publish and Subscribe (pub/sub) system is an established paradigm to disseminate the data from publishers to subscribers in a loosely coupled manner using a network of dedicated brokers. However, sensitive data could be exposed to malicious entities if brokers get compromised or hacked; or even worse, if brokers themselves are curious to learn about the data. A viable mechanism to protect sensitive publications and subscriptions is to encrypt the data before it is disseminated through the brokers. State-of-the-art approaches allow brokers to perform encrypted matching without revealing publications and subscriptions. However, if malicious brokers collude with malicious subscribers or publishers, they can learn the interests of innocent subscribers, even when the interests are encrypted. In this article, we present a pub/sub system that ensures confidentiality of publications and subscriptions in the presence of untrusted brokers. Furthermore, our solution resists collusion attacks between untrusted brokers and malicious subscribers (or publishers). Finally, we have implemented a prototype of our solution to show its feasibility and efficiency.

Cloud-assisted IoT applications are gaining an expanding interest, such that IoT devices are deployed in different distributed environments to collect and outsource sensed data to remote servers for further processing and sharing among users. On the one hand, in several applications, collected data are extremely sensitive and need to be protected before outsourcing. Generally, encryption techniques are applied at the data producer side to protect data from adversaries as well as curious cloud provider. On the other hand, sharing data among users requires fine grained access control mechanisms. To ensure both requirements, Attribute Based Encryption (ABE) has been widely applied to ensure encrypted access control to outsourced data. Although, ABE ensures fine grained access control and data confidentiality, updates of used access policies after encryption and outsourcing of data remains an open challenge. In this paper, we design PU-ABE, a new variant of key policy attribute based encryption supporting efficient access policy update that captures attributes addition to access policies. PU-ABE contributions are multifold. First, access policies involved in the encryption can be updated without requiring sharing secret keys between the cloud server and the data owners neither re-encrypting data. Second, PU-ABE ensures privacy preserving and fine grained access control to outsourced data. Third, ciphertexts received by the end-user are constant sized and independent from the number of attributes used in the access policy which affords low communication and storage costs.

Outsourcing sensitive data and operations to untrusted cloud providers is considered a challenging issue. To perform a search operation, even if both the data and the query are encrypted, attackers still can learn which data locations match the query and what results are returned to the user. This kind of leakage is referred to as data access pattern. Indeed, using access pattern leakage, attackers can easily infer the content of the data and the query. Oblivious RAM (ORAM), Fully Homomorphic Encryption (FHE), and secure Multi- Party Computation (MPC) offer a higher level of security but incur high computation and communication overheads. One promising practical approach to process the outsourced data efficiently and securely is leveraging trusted hardware like Intel SGX. Recently, several SGX- based solutions have been proposed in the literature. However, those solutions suffer from side channel attacks, high overheads of context switching, or limited SGX memory. In this paper, we present an SGX-assisted scheme for performing search over encrypted data. Our solution protects access pattern against side channel attacks while ensuring search efficiency. It can process large databases without requiring any long-term storage on SGX. We have implemented a prototype of the scheme and evaluated its performance using a dataset of 1 million records. The equality query and range query can be completed in 11 and 40 milliseconds, respectively. Comparing with ORAM- based solutions, such as ObliDB, our scheme is more than 10x faster.

In this paper, we propose a novel constant-size threshold attribute-based signcryption scheme for securely \nsharing data through public clouds. Our proposal has several advantages. First, it provides flexible cryptographic access control, while preserving users’ privacy as the identifying information for satisfying the access \ncontrol policy are not revealed. Second, the proposed scheme guarantees both data origin authentication and \nanonymity thanks to the novel use of attribute based signcryption mechanism, while ensuring the unlinkability \nbetween the different access sessions. Third, the proposed signcryption scheme has efficient computation cost \nand constant communication overhead whatever the number of involved attributes. Finally, our scheme satisfies strong security properties in the random oracle model, namely Indistinguishability against the Adaptive \nChosen Ciphertext Attacks (IND-CCA2), Existential Unforgeability against Chosen Message Attacks (EUFCMA) and privacy preservation of the attributes involved in the signcryption process, based on the assumption \nthat the augmented Multi-Sequence of Exponents Decisional Diffie-Hellman (aMSE-DDH) problem and the \nComputational Diffie Hellman Assumption (CDH) are hard.

International audience