Cloud platforms have become a primary environment for training machine learning models. However, the security of sensitive data could be compromised due to storage and computation in the cloud provided by the third party. Fully Homomorphic Encryption over the Torus (TFHE) offers a promising solution to this issue. However, integrating the wildly used Artificial Neural Networks (ANNs) with TFHE still faces significant challenges, particularly in the efficient implementation of activation functions and arithmetic operations. In this work, we present a parallel implementation of activation functions and arithmetic operations based on TFHE. By leveraging the parallel computing capabilities of multi-core CPUs and GPUs, our approach significantly improves computational performance. Experimental results demonstrate that our method achieves competitive accuracy while delivering a 12x speedup for 32-bit multiplication on GPUs and up to 30x acceleration for ReLU computations across various bit widths.
The Internet of Things (IoT) is fundamentally reshaping various industries by enabling efficient data collection and sharing. However, the prevalence of centralized cloud architectures introduces significant latency and high computational burdens, which are often incompatible with the real-time demands of resource-constrained edge devices. While edge computing provides a viable alternative, critical issues of data privacy and integrity remain, as a single data breach can lead to severe consequences. To address this dual challenge of security and efficiency, we introduce a novel edge-based computing framework. Our approach leverages secret-sharing-based Secure Multi-Party Computation alongside a terminal data source auditing mechanism to achieve secure model inference. We developed highly optimized protocols for Neural Network inference layers, significantly reducing communication overhead by integrating both arithmetic and boolean secret sharing. An innovative secure comparison protocol further minimizes server interactions, thereby boosting the efficiency of nonlinear computations. In addition, our framework incorporates a signature scheme to verify data legitimacy. Experimental evaluations on public datasets confirm that our method surpasses existing solutions in overall performance and security, validating its practical feasibility in modern IoT environments.
Federated learning (FL) enables social networks to utilize distributed user data without direct exposure. However, when FL updates are encrypted to prevent gradient inversion and privacy leakage, observability is significantly reduced and traditional anomaly detection is weakened. This creates opportunities for ciphertext poisoning while cryptographic audits incur prohibitive costs at scale. The key challenge is to design defenses that balance privacy protection, detection capability, and resource constraints. In this paper, we address this challenge by proposing an evolutionary game theoretic defense framework for encrypted FL in social networks. The framework models detection intensity of the server and poisoning propensity of clients as co-evolving strategies. It explicitly incorporates false positive penalties, external incentives for attackers, and practical constraints such as bandwidth and energy budgets. Through equilibrium analysis, we show how detection policies adapt to varying costs and incentives, and we identify stable strategy profiles under different deployment conditions. The framework yields actionable guidance for tuning server-side defenses, offering adaptive and cost-aware protection against ciphertext poisoning in large-scale social networks.
With rapid development of smart grids, how to efficiently aggregate and verify multi-source electricity consumption data while ensuring user privacy has become a key research problem, and a number of privacy-preserving data aggregation schemes have been introduced to alleviate this problem. However, most existing schemes only support simple additive homomorphic aggregation, lacking support for more complex functions. To address this challenge, this paper proposes an efficient and verifiable privacy-preserving data aggregation scheme that integrates functional encryption, Diffie–Hellman key exchange, and Schnorr signature to support linear function queries, resists collusion attacks, and ensures verifiability of aggregation results. Experimental results demonstrate that our proposed scheme achieves higher computational and communication efficiency compared with the existing schemes, e.g., GuardGrid, while maintaining strong security and functional expressiveness. Therefore, it provides a feasible and effective approach for building efficient, scalable, and verifiable privacy-preserving systems in smart grids.
With the growing demand for cross-platform collaboration in blockchain systems, cross-chain interoperability mechanisms have attracted widespread attention. However, the lack of trust among cross-chain entities poses serious challenges to the privacy protection of inter-chain data. Attribute-based access control (ABAC) enables fine-grained and flexible permission management, but existing schemes often disclose policies and attributes in plaintext, making it easier for attackers to exploit public on-chain information to launch attacks. To address this issue, this paper proposes a cross-chain access control scheme that combines ABAC with threshold fully homomorphic encryption (ThFHE). The scheme utilizes Cheon–Kim–Kim–Song (CKKS), a widely-used fully homomorphic encryption (FHE) scheme, to encrypt access policies and user attributes. Benefiting from the properties of FHE, smart contracts can perform permission verification directly in the ciphertext domain without exposing any sensitive information. Furthermore, the scheme achieves decentralized collaborative decryption by aggregating the partial decryption results from multiple relay nodes. Experimental results demonstrate that the scheme achieves secure and efficient cross-chain data access control with reasonable computational overhead.
In real-world applications where massive datasets are often stored across distributed computing units, traditional centralized semi-supervised learning (SSL) methods face significant challenges. These limitations include potential privacy risks and sensitivity to non-Gaussian noise. To address these issues, this paper proposes a novel distributed SSL approach that enhances privacy protection by processing data locally and eliminating the need for raw data exchange between units. Unlike conventional SSL algorithms rooted in the minimum mean squared error criterion, our method adopts the maximum correntropy criterion, substantially improving predictive accuracy in the presence of non-Gaussian noise frequently encountered in real-world data. Experiments on both synthetic and real-world datasets confirm that the proposed method outperforms existing approaches in terms of both prediction accuracy under non-Gaussian noise and privacy preservation.
Research focuses on the physical layer security (PLS) of mobile antenna arrays (MA) in near-field (NF) scenarios. A base station (BS) with an MA array serves a user with a single antenna, while multiple eavesdroppers with single antennas listen in. In order to reduce the risk of eavesdropping, a near-field MA-assisted secure transmission framework is proposed. The aim of this framework is to maximize achievable secrecy rates by jointly optimizing the base station’s beamforming vectors and mobile antenna positions. For this problem, high-quality suboptimal solutions are obtained through the Alternating Optimization (AO) algorithm and the Projected Gradient Ascent (PGA) method. Simulation results demonstrate that, compared to secure communication assisted by far-field MA and secure communication assisted by fixed antennas (FPA) in the near-field, this near-field MA secure transmission framework significantly enhances the security of information transmission. Specifically, when the base station transmission power is 10 W, the secure transmission rate increases by approximately 15 % and 18 % , respectively.
In recent years, satellite terrestrial integrated networks (STIN) have become crucial global communication infrastructure, playing irreplaceable roles in emergency response. However, their open architecture and dynamic nature expose them to security threats like identity spoofing and data eavesdropping. While current solutions have flaws, including single-factor authentication vulnerabilities and centralized architecture risks. In this paper, we propose a lightweight authentication protocol featuring security and dynamic key management. It integrates Physical Unclonable Functions (PUF), biometrics, and lightweight cryptography with hash-chain-based key updates. Formal verification under the Random Oracle Model (ROM) confirms the protocol resists ephemeral secret leakage and machine learning attacks while balancing anonymity and traceability. Compared to existing schemes, it maintains strong security while significantly reducing computational overhead, offering an efficient solution for STIN authentication.
To realize a chaotic system in digital world always leads a dynamical degradation of chaotic characteristics. Digital chaotic systems are increasingly utilized in modern cloud computing security architectures and big data analytics frameworks where precision degradation in large-scale, distributed, and heterogeneous environments can severely affect reliability. This paper focuses on the effects of fluctuations on digital chaotic systems and dynamical degradation related issues of those systems. A scheme based on hybrid structure is proposed for the solution to chaos degradation, in which the continuous chaotic system provides intrinsic random fluctuation. Further, the relationship between scaling of intrinsic randomness and chaos with different structure is studied. Symbolic dynamics is applied to rigorously prove that a class of digital systems is chaotic in the presence of fluctuations and to show that scaling behavior of fluctuations determines topological structure of chaos. Specifically, there are two critical fluctuation amplitudes at which topological structure of chaos essentially changes. Moreover, simulation studies are conducted to verify the theoretical results and further compare the effects of different random fluctuations on chaos degradation. These findings contribute to chaos-preserving mechanism designs applicable to precision-sensitive large-scale computing scenarios.
Generative steganography embeds secret messages during the image synthesis process, producing artificial images that lack the distortion signatures exploited by classical steganalysis. This shift introduces substantial challenges for reliable detection, particularly when the detector encounters stego images generated by previously unseen models. To address these limitations, we propose WaReCo, a wavelet residue contrastive learning framework designed to improve both feature sensitivity and cross-model generalization in deep steganalysis. Our method employs multi-level discrete wavelet transform to decompose images into frequency subbands, capturing high-frequency residual anomalies introduced by embedding operations. Supervised contrastive learning enhances feature discriminability by maximizing intra-class similarity and inter-class separation without requiring traditional cover-stego pairs. Extensive experiments on typical generative steganography methods across multiple datasets demonstrate that WaReCo consistently outperforms existing steganalysis networks. In particular, it achieves substantial gains in cross-model detection accuracy, often improving performance by more than ten percentage points over prior methods when training and testing generative models are different, validating WaReCo’s practical applicability for real-world generative steganalysis.
With the increasing reliance on digital identity, Decentralized Identity has become a promising paradigm to enhance user autonomy and privacy. However, existing decentralized identities face challenges such as secure key management, device migration, and unlinkability across contexts. To address these issues, we propose PMDID, a privacy-preserving and migratable decentralized identity system. First, PMDID introduces a two-layer identity mechanism in which a master ID derived from the ePassport (eID) enables Sybil resistance, while context-specific identities are registered via zero-knowledge proofs of master ID possession to ensure unlinkability across scenarios. Second, PMDID provides secure key management by combining Physical Unclonable Functions (PUFs) and biometric features with error correction and key derivation, guaranteeing key uniqueness, non-clonability, and recoverability. Third, PMDID designs a user registration protocol that supports secure identity migration, where a Merkle-structured issuance list maintained on the blockchain allows users to restore identity control on new devices without exposing sensitive information. We further implement a prototype and evaluate its performance, demonstrating that PMDID achieves strong guarantees of security, privacy, and scalability.
Satellite-Terrestrial Integrated Networks (STIN) hold significant promise for future global communication, but robust authentication remains a primary challenge for their secure application. Current authentication schemes often exhibit deficiencies such as inadequate dynamic binding of device biometrics, insufficient lightweight computation for resource-constrained satellites, lack of truly decentralized trust, and inefficient group handover in high concurrency scenarios. To address these issues, this paper proposes FG-LAS, a lightweight, fault-tolerant, and group handover authentication scheme. FG-LAS innovatively: 1) Introduces a Device Biometric Binding Key (DBBK) to achieve dynamic, unique, and replay-resistant binding between user biometrics and device fingerprints; 2) Employs Threshold ECDSA to construct a decentralized trust architecture, enhancing key management robustness and mitigating single-point-of-failure risks; and 3) Leverages satellite orbit predictability to pre-generate authentication parameters and optimize handovers using group processing, thereby reducing latency and computational overhead. Security analysis and performance evaluations demonstrate the proposed scheme’s feasibility, efficiency, and effectiveness for STIN environments.
Is there a secure way to share cyber threats knowledge among multiple organisations for a collective defense strategy ? Federated Learning (FL) has been introduced to enable collaboration among multiple organizations. In this paper, we propose a privacy-preserving federated learning approach to share cyber threat intelligence. The distributed nature of FL gives rise to different threats, such as inference attacks, poisoning attacks, and identity theft. In this work, we consider inference attacks, where a malicious subset of participant nodes aim to infer the training data of the victim. The attacker may generate and optimize vectors of features values to infer these data. We suggest a Federated Learning Approach Against Inference Attacks and Risks (FLAIR) where nodes collaboratively train models without a centralized server. We measured the data generated by the attacker against the real data of the target victim, and we show that the accuracy of this inference attack is low. We additionally conducted a set of experiments to test the performance and resilience of FLAIR. We then attended a high performance metrics with an accuracy around eighties.
This research investigates user perceptions of security and privacy in the context of AI technologies, focusing specifically on ChatGPT. We conducted a sentiment analysis of approximately 11,000 tweets collected from X Platform (formerly Twitter) between November 2022 and January 2024. Advanced natural language processing techniques were employed to preprocess the tweets, eliminating irrelevant data and refining the text for analysis. Sentiment analysis tools were then used to categorize user sentiments as positive, negative, or neutral. The results reveal a complex duality in user attitudes towards ChatGPT. While users generally appreciate the functional benefits and innovative capabilities of ChatGPT, there is substantial concern regarding data privacy and the potential misuse of AI technologies. Positive sentiments often highlighted the efficiency and utility of ChatGPT in various applications, whereas negative sentiments were predominantly focused on privacy risks and ethical considerations. These findings underscore the need for developers to prioritize transparent data handling practices and incorporate robust security features to address user concerns. Additionally, the study highlights the importance of updated regulations that protect user data while fostering innovation. Policymakers are urged to develop comprehensive policies that balance the need for security with the benefits of technological advancement. This study contributes to a deeper understanding of user perceptions of AI technologies, emphasizing the importance of a balanced approach that considers both benefits and risks. The insights gained provide a foundation for future research and inform strategies to enhance user trust and acceptance of AI technologies like ChatGPT.
This paper evaluates a local differential privacy (LDP) approach designed to address the challenges inherent in smartwatch datasets, particularly those dominated by numerical data. Traditional methods such as Lopub and Locop have demonstrated limitations in accurately estimating joint probability distributions (JPD) within these contexts. The Castell2D approach leverages a refined mechanism for perturbation and aggregation that enhances the precision of numerical data handling while maintaining robust privacy guarantees. In smartwatch datasets, the predominance of numerical data poses significant hurdles for existing LDP techniques, which often work with categorical data but fail with continuous values. By incorporating statistical techniques and noise mechanisms, Castell2D achieves a more accurate estimation of JPD. Comparative analyses with established approaches such as Lopub, Locop, and others show that Castell2D consistently outperforms them in terms of estimation accuracy and privacy preservation. The effectiveness of Castell2D is demonstrated through extensive experiments on real-world open smartwatch datasets from three of the largest companies-Fitbit, Apple Watch, and Garmin-and demonstrates its ability to maintain high utility of the data while ensuring privacy standards. This work represents a significant advancement in the field of LDP, providing a practical and effective solution for privacy-preserving data analysis in environments dominated by numerical data.
Social media influencers, those with verified accounts or with more than 10,000 followers, played a crucial role in the propagation of narratives during the COVID-19 pandemic. We investigate their impact by characterizing and contrasting the differences in content patterns between influential individuals versus public organizations during the pandemic, analyzing emotions, sentiments, and scientific claims expressed in their Tweets. Advanced machine learning approaches, including customized transformer models, few-shot learning, and large language models such as GPT-3.5, were used. The findings reveal a stark contrast in sentiment usage across sub-domains like vaccines and lockdowns, with organizations predominantly employing neutral tones while individuals displaying a significant negative sentiment bias. Individuals often conveyed more negative emotions, whereas organizations exhibited greater optimism. However, many claims from both groups were not verified, highlighting the need to combat misinformation.
As seniors continue to age and look to age in place by utilizing smart devices with sensors, online platforms are a major resource for them and their care givers to socialize and search or look for information on agetech. This comes with a risk as they can be unfairly targeted or influenced in a certain way. Influence refers to information or action that can cause a change in direction, thought or action either positively or negatively, some sort of coordinated manipulation either positively or negatively. There are many influence campaigns propagated through social media and our focus is on information that targets the elderly particularly on the topic of agetech. This paper presents a review of existing work around influence campaigns in general, the methods used in influence detection and the obtained results. To our knowledge none of the existing work has addressed influence campaigns from agetech perspective. With that in mind, we developed a new approach for detecting influence campaigns aimed at agetech using machine learning and classifier calibration. We used machine learning to highlight influence campaign and misinformation in agetech. To develop and evaluate our approach, we collected an agetech dataset consisting of a set of tweets that contain the hashtag agetech. The evaluation yielded encouraging results.
This paper examines the application of blockchain technology in electric vehicle (EV) charging payment systems, focusing on a comparative analysis of four EVM-supported platforms: Binance Smart Chain, Polygon, Fantom, and Celo. As the EV market expands, traditional payment methods struggle to meet the unique requirements of charging infrastructure, such as real-time pricing and seamless user experiences across networks. We analyze these platforms' performance in three key operations: transaction creation, NFT minting, and NFT transfer, considering factors like transaction fees, gas limits, gas usage, and gas prices. The study aims to provide insights into the suitability of each platform for real-world EV charging applications, addressing the challenges of implementing blockchain-based payment in this domain.
Recent advances in artificial intelligence have prompted the use of machine learning methods in network security. In this paper, we address the issue of imbalanced data that is often present in network security datasets used in machine learning. We propose an oversampling method based on the Gamma distribution to balance the data prior to training. The results on several imbalanced datasets show the potential of the proposed method as a viable tool to build intrusion detection systems based on artificial intelligence. The accompanying code for the study is available on Github.
The majority of Australian children aged between 14–18 have a smartphone. Children at this age are concerned more with sharing information with their friends than with protecting their privacy. For this reason, they tend to leave their smartphone's default privacy settings on making it possible for anyone to discover their movements. This project aimed at raising children’s awareness about the security issues associated with their smartphones. M5Stack Core 2 ESP32 devices programmed to detect Bluetooth and WI-FI signals were installed around two high school campuses in Australia to capture traffic within their range. Children were then informed of the presence of these network sensors in their campuses, during a cyber safety lesson in which they were shown how to modify their smartphone privacy settings to prevent their smartphones from sharing information about them without their knowledge. Samples of data extracted from the network sensors before and after the delivery of the cyber safety lesson were compared to find out if the children in these two schools changed their smartphone settings to stop their sensitive information from escaping from their smartphones. The outcome of the analysis of the before and after data showed that the volume of traffic captured from smartphones after the cyber safety lesson was significantly less than the volume of traffic captured before the cyber safety lesson, suggesting children must have benefited from the lesson.
