Integrating Large Language Models (LLMs) into robotic systems has revolutionised embodied artificial intelligence, enabling advanced decision-making and adaptability. However, ensuring reliability — encompassing both security against adversarial attacks and safety in complex environments — remains a critical challenge. To address this, we propose a unified framework that mitigates prompt injection attacks while enforcing operational safety through robust validation mechanisms. Our approach combines prompt assembling, state management, and safety validation, evaluated using both performance and security metrics. Experiments show a 30.8% improvement under injection attacks and up to a 325% improvement in complex environment settings under adversarial conditions compared to baseline scenarios. This work bridges the gap between safety and security in LLM-based robotic systems, offering actionable insights for deploying reliable LLM-integrated mobile robots in real-world settings. The framework is open-sourced with simulation and physical deployment demos at https://llmeyesim.vercel.app/ .

Federated learning (FL) has been widely adopted to provide machine learning (ML) privacy, protecting sensitive user data from leakage. However, there are still attacks that could exploit FL to access users' sensitive data, such as model inversion attacks, property inference attacks, and membership inference attacks. Various solutions were proposed to secure FL using various privacy-preserving techniques, such as differential privacy, homomorphic encryption, and multi-party encryption. However, existing solutions often add noise to the model that hinders the accuracy, or introduce large computational overhead that makes them impractical to use. In this paper, we propose a new privacy protection scheme for FL that uses homomorphic encryption (HE), noise, and secret sharing to protect users' sensitive data from up to n-2 adversarial clients and the server colluding. The computational overhead is minimised by transferring expensive computations of HE to the server, requiring only the encryption and homomorphic addition to be carried out by clients. We provide proof sketches to validate the security of our scheme, and experimental results to demonstrate the practicality of our proposed scheme. The results show that our scheme adds only up to 8% overhead without losing any accuracy to base FL models, showing minimal overhead without losing accuracy, regardless of the data used.

Large language models (LLMs) have emerged as a promising tool for detecting code vulnerabilities, potentially offering advantages over traditional rule-based methods. This paper proposes an enhanced framework for vulnerability detection using LLMs, incorporating various prompt engineering strategies to improve performance. We evaluate several techniques, including role-based prompting, zero-shot chain-of-thought, and structured prompting approaches, on the DiverseVul dataset of C/C++ vulnerabilities. Our experiments assess the framework’s performance across different code structures, contextual information levels, and LLM capabilities. Our results show that using our dynamic prompt engineering technique, you can improve the F1 score by up to 100% with GPT-3.5, a widely used LLM model. We also observe that GPT-4o, Gemini 2.0 Flash, and Meta Llama 3.1 generally outperform GPT-3.5, and all models are very poor when it comes to correctly identifying the type of vulnerability in the code, with the best F1 score of 0.16 observed. However, our follow-up experiments on LLM-based vulnerability correction (i.e., patching) show a 45.77% success rate using GPT-4o, demonstrating promising results in leveraging LLMs for enhancing software security and providing insights into optimizing prompt engineering for vulnerability detection tasks.

In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities in understanding and solving mathematical problems, leading to advancements in various fields. We propose an LLM-embodied path planning framework for mobile agents, focusing on solving high-level coverage path planning issues and low-level control. Our proposed multi-layer architecture uses prompted LLMs in the path planning phase and integrates them with the mobile agents' low-level actuators. To evaluate the performance of various LLMs, we propose a coverage-weighted path planning metric to assess the performance of the embodied models. Our experiments show that the proposed framework improves LLMs' spatial inference abilities. We demonstrate that the proposed multi-layer framework significantly enhances the efficiency and accuracy of these tasks by leveraging the natural language understanding and generative capabilities of LLMs. Experiments conducted in our Eye Sim simulation demonstrate that this framework enhances LLMs' 2D plane reasoning abilities and enables the completion of coverage path planning tasks. We also tested three LLM kernels: gpt-4o, gemini-1.5-flash, and claude-3.5-sonnet. The experimental results show that claude-3.5 can complete the coverage planning task in different scenarios, and its indicators are better than those of the other models. We have made our experimental simulation platform, Eye Sim, freely available at https://roblab.org/eyesim/.

Smart metering systems (SMS) are popular in industrial and residential areas but can risk privacy by revealing user behaviors. Homomorphic encryption (HE) is a technique that protects data privacy by enabling calculations on encrypted data. However, the high computational costs of HE can hinder real-time or resource-constrained applications. Our paper presents a framework to encrypt only selected SMS data for improved efficiency without compromising privacy significantly. By encrypting data blocks with higher entropy values, we can mitigate the leakage of key information to adversaries who may conduct privacy attacks, such as membership inference attacks (MIA). We evaluate our framework using two real-world datasets (i.e., electricity and water) to assess privacy and performance trade-offs. The results indicate a 47% performance increase while still providing a sufficient level of privacy when adopting an encryption ratio of 0.4. They demonstrate the effectiveness of the proposed framework, considering the trade-off between privacy and performance, where the user can determine the appropriate security level for privacy protection and enhance the performance using HE in practical SMS settings.

Spam filtering services typically operate via cloud outsourcing, which exposes sensitive and private email content to the cloud server spam filter. Homomorphic encryption (HE) can address this issue by ensuring that user emails remain encrypted throughout all stages of the spam detection process on the cloud server. However, existing HE-based approaches are computationally infeasible due to the nature of HE operations. This paper proposes BlindFilter, a distributed, lightweight, HE-based spam email detection approach that consists of clients and servers collaborating to perform spam detection operations securely. BlindFilter employs WordPiece encoding and a modified Naive Bayes classifier, mitigating the need for multiplications and comparisons that would be prohibitive in terms of computation when applied with HE. Our experimental results demonstrate the efficacy of BlindFilter, with F1 scores exceeding 97% across two public email datasets. Furthermore, BlindFilter proves to be efficient as it can process an email in an average of 482.78 milliseconds. Our analysis also reveals that BlindFilter is robust against model extraction attacks, in which malicious users attempt to deduce the features of BlindFilter from query-response pairs.

The satisfaction of a software requirement is commonly stated as a Boolean value, that is, a security requirement is either satisfied (true) or not (false). However, a discrete Boolean value to measure the satisfaction level of a security requirement by deployed mechanisms is not very useful. Rather, it would be more effective if we could quantify the level of satisfaction of security requirements on a continuous scale. We propose an approach to achieve this for cloud software systems based on relationships between defense strength, exploitability of vulnerabilities, and attack severity. We extend the concept of entailment relationship from the field of requirements engineering with the satisfiability aspects of security requirements. The proposed approach enables us to systematically structure security concepts into three sets of related descriptions to quantify the satisfaction level of security requirements with the deployed security solutions. To demonstrate the feasibility of the proposed approach, we evaluate the approach in a case study. As a result, security administrators are able to deploy more effective and appropriate security solutions based on their assessment.

It is still challenging to effectively identify rumors due to rapid changes in people's interests and perceptions. To enhance rumor detectors, we first need to better understand which rumors are effective (in terms of bypassing detection) and their characteristics. In this paper, we introduce ARGH, a novel framework to automatically generate rumors using recent advancements in natural language processing, customized to target and generate specific topics. To show the effectiveness of ARGH, we conducted a user study with 212 participants and analyzed how well humans can detect the rumors generated by ARGH, and we also tested its performance against the state-of-the-art rumor detection model PLAN [17]. Surprisingly, the experimental results demonstrate that the generated rumors are significantly harder to identify as rumors than hand-written rumors, degrading the detection accuracy by both humans and machines by 18.87% and 17.62%, respectively. We believe that ARGH will be a useful tool to obtain high quality and evasive rumor datasets quickly, which is often a tedious and time consuming task. Further, our analysis results provide valuable insight into how to characterize evasive rumors and how they can be generated, which will help to enhance the existing rumor detection techniques.

Existing security risk evaluation approaches (e.g., asset-based) do not consider specific security requirements of individual cloud computing clients in the security risk evaluation. In this paper, we propose a threat-specific risk evaluation approach that uses various security attributes of the cloud (e.g., vulnerability information, the probability of an attack, and the impact of each attack associated with the identified threat(s)) as well as the client-specific security requirements in the cloud. Our approach allows a security administrator of the cloud provider to make fine-grained decisions for selecting mitigation strategies in order to protect the outsourced computing assets of individual clients based on their specific security needs against specific threats. This is different from the existing asset-based approaches where they do not have the functionalities to provide the security evaluation of the cloud with respect to specific threats. On the other hand, the proposed approach enables security administrators to compute a range of more effective client-specific countermeasures with respect to the importance of security requirements and threats. The experimental evaluation results demonstrate that effective security solutions vary due to specific threats prioritized by different clients for an application in the cloud. Further, the proposed approach is not limited to only the cloud-based systems, but can easily be adopted to other networked systems. We have also developed a software tool to support the proposed approach.

Software defined networking (SDN) has been adopted in many application domains as it provides functionalities to dynamically control the network flow more robust and more economical compared to the traditional networks. In order to strengthen the security of the SDN against cyber attacks, many security solutions have been proposed. However, those solutions need to be compared in order to optimize the security of the SDN. To assess and evaluate the security of the SDN systematically, one can use graphical security models (e.g., attack graphs and attack trees). However, it is difficult to provide defense against an attack in real time due to their high computational complexity. In this paper, we propose a real-time intrusion response in SDN using precomputation to estimate the likelihood of future attack paths from an ongoing attack. We also take into account various SDN components to conduct a security assessment, which were not available when addressing only the components of an existing network. Our experimental analysis shows that we are able to estimate possible attack paths of an ongoing attack to mitigate it in real time, as well as showing the security metrics that depend on the flow table, including the SDN component. Hence, the proposed approach can be used to provide effective real-time mitigation solutions for securing SDN.

Hardening the dynamic networks is a very challenging task due to their complexity and dynamicity. Moreover, there may be multi-objectives to satisfy, while containing the solutions within the constraints (e.g., fixed budget, availability of countermeasures, performance degradation, non-patchable vulnerabilities, etc). In this paper, we propose a systematic approach to optimise the selection of the security hardening options for the dynamic networks given multiple constraints and objectives. To do so, we evaluate potential attack scenarios for a given time period, and then use a multi-objective optimisation based on Non-dominated Sorting Genetic Algorithm to find the optimal set of security hardening options. We measure the effectiveness of the options using various security metrics, which is demonstrated through experimental analysis. The results show that our approach can be applied to select the optimal set of security hardening options to be deployed for the dynamic networks given multiple objectives and constraints.

There are advances in detecting malware using machine learning (ML), but it is still a challenging task to detect advanced malware variants (e.g., polymorphic and metamorphic variations). To detect such variants, we first need to understand the methods used to generate them to bypass the detection methods. In this paper, we introduce an adaptive malware variant generation (AMVG) framework to study bypassing malware detection methods efficiently. The AMVG framework uses ML (e.g., genetic algorithm (GA)) to generate malware variants that satisfy specific detection criteria. The use of GA automates the malware variant generations with appropriate modules to handle various input formats. For the experiment, we use malware samples retrieved from theZoo, a collection of malware samples. The results show that we can automatically generate malware variants that satisfy varying detection criteria in a practical amount of time, as well as showing the capabilities to handle different input formats.

The Internet of Things (IoT) is a network comprised of heterogeneous devices that can exchange data without requiring human-to-human or human-to-computer interactions. However, there are various vulnerabilities found due to the heterogeneity of the IoT network. Moreover, the mobility of IoT devices causes potential dynamic changes to the attack surfaces of IoT networks. As a result, static network security analysis approaches cannot capture these changes. In order to address this problem, we present an IoT security assessment approach by modelling different movement patterns of mobile IoT devices. Graphical security models are used in conjunction to evaluate the security of the IoT networks taking into account the mobility of the IoT devices. Further, we use various security metrics to analyze the security of the network to show the changing security posture when mobility is taken into account. The feasibility of the proposed approach is demonstrated by analyzing the security of an example mobile IoT network using three existing synthetic mobility models: Random Waypoint, Gauss-Markov and Reference Point Group. The experimental analysis shows the changing attack surface of the IoT networks when mobile devices are considered.

Conventional security risk assessment approaches for cloud infrastructures do not explicitly consider risk with respect to specific threats. This is a challenge for a cloud provider because it may apply the same risk assessment approach in assessing the risk of all of its clients. In practice, the threats faced by each client may vary depending on their security requirements. The cloud provider may also apply generic mitigation strategies that are not guaranteed to be effective in thwarting specific threats for different clients. This paper proposes a threat-specific risk assessment framework which evaluates the risk with respect to specific threats by considering only those threats that are relevant to a particular cloud client. The risk assessment process is divided into three phases which have inter-related activities arranged in a spiral. Application of the framework to a cloud deployment case study shows that considering risk with respect to specific threats leads to a more accurate quantification of security risk. Although our framework is motivated by risk assessment challenges in the cloud it can be applied in any network environment.

A Software Defined Network (SDN) provides functionalities for modifying network configurations. To enhance security, Moving Target Defense (MTD) techniques are deployed in the networks to continuously change the attack surface. In this paper, we realize an MTD system by exploiting the SDN functionality to optimally reconfigure the network topology. We introduce a novel problem Shuffle Assignment Problem (SAP), the reconfiguration of a network topology for enhanced security, and we show how to compute the optimal solution for small-sized networks and the near-optimal solution for large-sized networks using a heuristic method. In addition, we propose a shuffle-based online MTD mechanism, which periodically reconfigures the network topology to continuously change the attack surface. This mechanism also selects an optimal countermeasure using our proposed topological distance metric in real-time when an attack is detected. We demonstrate the feasibility and the effectiveness of our proposed solutions through experimental analysis on an SDN testbed and simulations.

Cyber crime is a developing concern, where criminals are targeting valuable assets and critical infrastructures within networked systems, causing a severe socio-economic impact on enterprises and individuals. Adopting moving target defense (MTD) helps thwart cyber attacks by continuously changing the attack surface. There are numerous MTD techniques proposed in various domains (e.g., virtualized network, wireless sensor network), but there is still a lack of methods to assess and compare the effectiveness of them. Security models, such as an attack graph (AG), provide a formal method of analyzing the security, but incorporating MTD techniques in those security models has not been studied. In this paper, we incorporate MTD techniques into a security model, namely a hierarchical attack representation model (HARM), to assess the effectiveness of them. In addition, we use importance measures (IMs) for deploying MTD techniques to enhance the scalability. Finally, we compare the scalability of AG and HARM when deploying MTD techniques, as well as changes in performance and security in our experiments.

Software Defined Network (SDN) is an emerging paradigm for flexible network design and implementation. Availability metric of SDNs is critically demanding further studies. This paper aims to propose hierarchical models to assess the availability of SDNs. We incorporate various failure modes and recovery behaviors in the SDN including (i) link failures at network level, and (ii) software and hardware failures at network device level. We use hierarchical models in which a Reliability Graph (RG) is used to represent the reachability of hosts (and switches) in the SDN at the upper level and Stochastic Reward Net (SRN)s are used to represent the detailed failure and recovery of network devices at the lower level, respectively. We incorporate the programmable capability of the SDN at the upper level (i.e., the RG). We perform numerical analysis to assess the availability of the SDN in terms of steady state availability and downtime in minutes per year, and we also show the sensitivity analysis.

With the rise in popularity of using websites to distribute content to users, content creators needed to gain revenue from page views. To achieve this, they use third-party networks to distribute advertisements and perform analytics on users to customize their advertisements. These networks build up a profile of their users (e.g., navigation patterns, visit frequency, personal interest), which poses a significant privacy risk as this information can be sold to others or used in targeted advertising. In order for users to protect themselves, they need to install add-ons for their web browser which removes third-party content. This paper surveys some of the existing add-ons to block third-party content, and determine which add-ons are required to provide adequate protection. Further analytics are conducted to evaluate the effectiveness of those add-ons through the experiments.

Moving Target Defense (MTD) changes the attack surface of a system that confuses intruders to thwart attacks. Various MTD techniques are developed to enhance the security of a networked system, but the effectiveness of these techniques is not well assessed. Security models (e.g., Attack Graphs (AGs)) provide formal methods of assessing security, but modeling the MTD techniques in security models has not been studied. In this paper, we incorporate the MTD techniques in security modeling and analysis using a scalable security model, namely Hierarchical Attack Representation Models (HARMs), to assess the effectiveness of the MTD techniques. In addition, we use importance measures (IMs) for scalable security analysis and deploying the MTD techniques in an effective manner. The performance comparison between the HARM and the AG is given. Also, we compare the performance of using the IMs and the exhaustive search method in simulations.

Computing a prioritized set of vulnerabilities to patch is important for system administrators to determine the order of vulnerabilities to be patched that are more critical to the network security. One way to assess and analyze security to find vulnerabilities to be patched is to use attack representation models (ARMs). However, security solutions using ARMs are optimized for only the current state of the networked system. Therefore, the ARM must reanalyze the network security, causing multiple iterations of the same task to obtain the prioritized set of vulnerabilities to patch. To address this problem, we propose to use importance measures to rank network hosts and vulnerabilities, then combine these measures to prioritize the order of vulnerabilities to be patched. We show that nearly equivalent prioritized set of vulnerabilities can be computed in comparison to an exhaustive search method in various network scenarios, while the performance of computing the set is dramatically improved, while equivalent solutions are computed in various network scenarios.