Privacy-Preserving Record-Linkage (PPRL) techniques provide an error-tolerant way of linking data records from different sources based on quasi-identifiers while keeping these identifying attributes private. Many practical use cases require non-interactive PPRL, where the data holders merely prepare their data and outsource data linkage to a third party. While many such schemes have been proposed and some are even used in practice, almost all of them are either insecure or impractical for various reasons. This work presents a novel PPRL scheme which utilizes a technique called ‘Polynomial Sparsity Testing’. We provide extensive theoretical and experimental evidence for the high linkage quality and security of the scheme, which is on par with the state of the art: Using comparable parameter choices in terms of security, our scheme achieves a F_1 -measure of 94.3% at a precision of 100% , while comparing a record pair takes less than 0.4 ms.
Symmetric functions are an important class of functions that includes many interesting functions. Constructing efficient card-based protocols for symmetric functions is considered an important research topic in card-based cryptography. In 2020, Ruangwises and Itoh designed a protocol for symmetric functions f: {0,1}^n → R using 2n + 2 cards of a two-colored deck. Whether the number of cards in this protocol could be reduced is an open problem. In this paper, we design protocols for symmetric functions using 2n + 1 cards of a two-colored deck, which solve the above problem affirmatively. We also propose protocols for symmetric functions using 2n cards of a three-colored deck, while the existing 2n-card protocol requires a four-colored deck. By further restricting the class of functions to {0} -partially doubly symmetric functions, we can construct even more efficient protocols. In particular, we construct protocols for {0} -partially doubly symmetric functions using 2n cards of a two-colored deck, while the existing protocol requires 2n+1 cards.
The rise of autonomous AI agents powered by large language models (LLMs) has been accompanied by new frameworks for integrating these agents with external tools and data. One such framework is Anthropic’s Model Context Protocol (MCP), a recently introduced open standard that enables AI assistants to connect with a wide variety of external systems. While MCP unlocks powerful capabilities for agentic AI, it also dramatically expands the supply chain threat surface. In this paper, we investigate supply chain threats in the MCP ecosystem, a rapidly emerging security frontier as community-driven development and open-source MCP servers become prevalent. We identify and categorize major supply chain threats in the MCP ecosystem and validate representative scenarios via proof-of-concept attacks. In particular, we demonstrate how malicious MCP servers, as well as hostile data inputs, can be used to intentionally trigger unauthorized or harmful behaviors, such as sensitive data exfiltration or security policy violations. Finally, we outline two complementary avenues for mitigating MCP supply-chain threats. First, a specification- and document-based validation framework addresses code-driven threats by statically verifying that an MCP tool’s implementation aligns with its declared interface and behavior. This method has been prototyped and shown to detect functional inconsistencies in real-world tools. Second, we propose the conceptual design of lightweight runtime validation agents—supervisory components that monitor prompt flows, tool responses, and runtime context to intercept data-driven threats. Together, these layers—proven static analyzers and envisioned runtime validators—form a cohesive foundation for securing AI-agent infrastructures in the MCP ecosystem. By identifying novel attack surfaces and proposing layered defenses, our work represents an early step toward framing supply chain threat dimensions in MCP and AI agent security, and contributes to the ongoing discourse on mitigation strategies for AI-integrated supply chains.
As online gaming and esports platforms grow in popularity and complexity, concerns over account security and user authentication are increasing. This study explores the use of behavioral biometrics to identify individual players based solely on their in-game actions in Counter-Strike: Global Offensive (CS:GO) using data from professional tournament matches, with a particular focus on the impact of sequence length on identification performance. We propose a machine learning framework that aggregates gameplay data into fixed length sequences and extracts interpretable features related to movement, aiming, and economic behavior. To investigate how temporal context influences model performance, we evaluate binary classifiers across a range of sequence durations, from a few seconds to several minutes. For each duration, behavioral data is aggregated into a single feature vector using statistical summaries, such as mean and count, to represent player behavior over the window. Importantly, the window sizes overlap every five seconds of gameplay, generating more data for the model and allowing for finer-grained analysis. To assess generalizability, models are tested exclusively on unseen matches occurring after the training period. Results show that identification accuracy is sensitive to sequence length, with longer sequences generally resulting in higher predictions. This case study contributes to the development of behavioral authentication systems in gaming by demonstrating how sequence length and feature design influence player identification performance.
Local Differential Privacy (LDP) has recently emerged as a popular standard for privacy-preserving data collection, and bitvector-based LDP protocols such as RAPPOR and OUE are widely used in both academic and industrial applications. To evaluate LDP protocols and applications, researchers commonly rely on simulation-based experiments, where multiple users’ perturbations are simulated sequentially on one computer. While faithful to protocol definitions, this approach incurs substantial execution times, especially for large user populations and domains. To address this concern and enable fast simulations, in this paper, we propose a novel simulation methodology for bitvector-based LDP protocols. Our key insight is to model the collective effect of randomized perturbation using Binomial random variables, avoiding the need to simulate each user individually. We theoretically and empirically show that this strategy reduces computational complexity while producing unbiased estimations with identical variance to RAPPOR and OUE. Furthermore, we empirically show that our method reduces execution times from several minutes to less than a second, yielding multiple orders of magnitude improvement. Overall, our work offers a fast and scalable method for simulating bitvector-based LDP protocols, with direct applicability to existing works and simulation platforms.
Infrastructure as Code (IaC) platforms lack mechanisms for detecting security smell diffusion, a challenge stemming from the absence of repository relationships. We present a similarity-based methodology combining content and structure metrics to identify repository clones. Validated against Ansible Galaxy repositories that have GitHub fork data, our approach achieved 99.6 × higher rates than platforms with visible fork relationships. A survey of 24 IaC tools confirmed none provide cross-repository comparison capabilities, demonstrating a gap in repository relationship tracking within the IaC supply chain. Our work addresses this gap by providing a systematic approach to detect clones and track security diffusion in environments lacking fork metadata.
Private Simultaneous Messages (PSM) is a kind of secure multiparty computation with minimal interaction pattern and minimal security requirement. A PSM protocol is said to be with universal reconstruction for a given function family if the algorithm of the referee (the output party) is independent of a function to be computed and the referee cannot infer the function from a protocol execution. In a recent work by Eriguchi and Shinagawa (EUROCRYPT 2025), the authors proposed a compiler to obtain a PSM protocol for symmetric functions from PSM protocols with universal reconstruction for symmetric functions with smaller domains. The authors also constructed the latter PSM protocols with universal reconstruction, by which the former PSM protocol achieves communication complexity better than the previously known protocols. In this paper, we construct the latter PSM protocols with universal reconstruction for symmetric functions more efficiently; the communication complexity is exponentially (in the input range) smaller than the protocols by Eriguchi and Shinagawa. As a consequence, we also obtain a PSM protocol (and also an ad-hoc PSM protocol and a robust PSM protocol) for symmetric functions that is more efficient than their protocol. Technically, a main ingredient of their protocols is a linear and injective encoding of histograms for the input elements, and our improvement is realized by finding a more efficient encoding of the histograms.
The importance of data-driven optimization in society and industry is increasing, including the utilization of data in enterprises and public institutions, as well as the advancement of Evidence-Based Policy Making (EBPM). However, conventional privacy protection technologies are becoming obsolete due to the evolution of attack techniques, making robust countermeasures an urgent necessity. To address this challenge, differential privacy (DP), a framework for protecting privacy that provides mathematical security against arbitrary attacks, has gained significant attention. In this paper, we focus on Zero-Concentrated Differential Privacy (zCDP), an extension of DP that incorporates Rènyi divergence into the security definition, and theoretically and experimentally evaluate the improvement of privacy protection for large-scale high-dimensional data. Specifically, we redefine the existing Non-Negative Wavelet (NN-Wavelet) method within the zCDP framework by replacing the Laplace mechanism with the Gaussian mechanism, and compare its performance with that of existing methods using population statistics based on census data. The evaluation results showed that the proposed method showed the highest accuracy for specific parameter settings. The composability of the proposed method in maintaining high accuracy across multiple data releases makes it especially effective for datasets such as demographic information, which are frequently updated and shared.
Proof-of-Concept (PoC) codes against vulnerabilities are widely available on various platforms including GitHub. Security researchers and vulnerability analysts can effectively use them for investigating vulnerabilities and creating attack signatures. However, various reports have warned the existence of fake PoC codes that aim to disguise themselves as legitimate PoC codes to make users execute malware like cryptojacking malware, infostealers, and bot malware. To prevent malware infections among users, it is desired to establish methods for automatically detecting fake PoC codes. One of existing studies has investigated fake PoC codes published on GitHub; however, its detection method for fake PoC codes produces many false positives because it only identifies potentially suspicious data like Base64-encoded strings using regular expressions without examining how such data is used within the code. In this paper, we design a novel detection method leveraging taint analysis, a type of static code analysis technique. Specifically, if a code contains potentially suspicious data, the data flows are inspected to detect code fragments that perform suspicious actions using the data, such as external communications or execution of obfuscated commands. By applying this method to PoC codes published on GitHub, we demonstrate that false positives can be reduced to less than 8.3
To reduce the spread of online disinformation by social media users, education to improve their knowledge and skills is being promoted. However, other factors contribute to the spread of disinformation, one of which is emotion, especially anger, and few measures have focused on this aspect. The study aim was to clarify the effect of nudges—which encourage users to deliberate by drawing their attention to the strong anger associated with disinformation—on reducing the spread. The effect of nudges was compared with that of education. We focused on nudges that use emotion regulation and measured responses in an experiment to test whether they reduced the sharing of disinformation more than education did. The results showed that nudges reduced the sharing of disinformation stimuli by 34.5
Managing noise growth is a central challenge in fully homomorphic encryption (FHE). Gadget decomposition mitigates this by representing elements as vectors whose inner product with a gadget vector approximately reconstructs the original value. Radix-based decompositions support approximation but CRT-based ones have, so far, required exactness. We introduce, for the first time, CRT-based gadget decompositions in the approximate setting, combining the benefits of approximate decompositions with the structural advantages of CRT-based methods. This enables efficient blind rotation and (programmable) bootstrapping in TFHE using only native arithmetic while increasing parallelism. On a typical FPGA (17-bit multipliers), our approach achieves a speedup of over 2× and approximately 50% lower area than comparable radix-based approximate designs. The methodology also reduces bandwidth, memory, and compute in settings with large ciphertext moduli (e.g., 128-bit), benefiting both hardware and software implementations.
Functional encryption (FE) enables fine-grained access control over encrypted data: each decryption key is tied to a function, allowing users to compute only the function’s output on the underlying message. While FE for general functions is hard to achieve, static bounded FE, where the adversary declares a fixed collusion bound q at the outset, can be constructed from minimal assumptions (e.g., public-key encryption (PKE) for public key FE and one-way functions (OWFs) for symmetric key FE). To enhance flexibility, Agrawal et al. (Crypto 2021) and Garg et al. (Eurocrypt 2022) introduced dynamic bounded public key FE, where the collusion bound q is chosen per encryption rather than fixed globally. In this paper, we study dynamic bounded FE in the symmetric key setting. We present a compiler that upgrades static bounded FE to dynamic bounded FE using pseudorandom functions (PRFs), preserving both function and message privacy. Unlike its public key counterpart, our compiler avoids the non-black-box tool of garbled circuits and relies solely on OWFs. Consequently, we show that dynamic bounded, function-private symmetric key FE can be built from OWFs alone.
Card-based protocols utilizing up-down cards with an asymmetrical mark on the face side, allowing it to be distinguished after a 180-degree rotation, have been studied. Previous research has proposed finite-time committed-AND protocols incorporating interlocking operations, which combine shuffling and random rotations of cards. These protocols achieve the theoretical minimum by using only three up-down cards. However, interlocking operations are difficult to implement in practice. In this paper, we prohibit interlocking operations and prove that the minimum number of cards required without such operations is four.
Multivariate public key cryptography (MPKC) is considered a promising candidate for post-quantum cryptography, with its security relying on the hardness of solving systems of multivariate quadratic equations. Among MPKC schemes, the unbalanced oil and vinegar (UOV) and its variants have been actively studied. Pébereau and Luyten showed that the Kipnis–Shamir attack and the singular point attack can be described within the same framework using the Jacobian matrix. In this study, we demonstrate that the rectangular MinRank attack can also be described within this framework. Furthermore, by leveraging this framework, we extend the feasible target ranks of the rectangular MinRank attack and use this extended attack to analyze the security of UOV and its variants. In conclusion, we confirm that the currently proposed parameters for UOV, MAYO, QR-UOV, and SNOVA are resistant to this attack.
The 2F construction was created in the hopes of defending known multivariate cryptosystems against rank-based attacks by employing modulus switching. This construction claimed post-quantum encryption with ciphertext sizes much smaller than prominent lattice-based schemes. However, the resulting structure inherent to any 2F cryptosystem yields new challenges in the form of lattice-based attacks. One attack in particular—the NTRU attack—greatly reduces the security of the 2F constructions. The resulting parameters to defend against this attack yield a cryptosystem with ciphertexts only marginally smaller than the lattice-based schemes. In this work, we investigate the security of a 2F cryptosystem, 2FSquare, against lattice-based attacks, determining the viability of the construction and finding realistic, secure parameters. As a point of comparison, we introduce and investigate rectangular variations on NTRU to determine the extent to which lattice-based schemes can achieve small ciphertexts and whether 2F holds an advantage over lattice-based encryption in this respect. We propose new parameter sets with several aspect ratios to determine the extent to which ciphertext size can be adjusted in NTRU-like schemes.
The accelerating growth in scale and complexity of modern processors has led to the constant identification of transient execution vulnerabilities (TEVs), predominantly found within x86-64 and ARM CPUs. The rise in adoption of the RISC-V architecture in recent years has coincided with a surge in TEV research of its implementations. Previous studies have showcased several TEV variants in RISC-V processors. However, the range of RISC-V processors employed as platforms for TEV research has been limited in diversity, and the Spectre variant Speculative Store Bypass (SSB) has not been evaluated in RISC-V processors that utilize a memory dependence predictor (MDP), a crucial feature of numerous out-of-order (OoO) processors across different instruction set architectures (ISAs) including RISC-V and others. In this paper, we begin by examining the SSB vulnerability in a 32-bit OoO MDP-equipped RISC-V processor, “RSD”, which is compact and efficient, possessing potential applications in alternative contexts in contrast to earlier research that concentrated on heavier 64-bit CPUs. Following this, we replicate the SSB attack gadget in RSD using the Verilator software simulator and a ZedBoard Zynq-7000 FPGA board. Subsequently, we utilize the Konata pipeline viewer to visualize and verify the results obtained, confirming that the SSB attack remains feasible with the partial defense offered by an MDP. Furthermore, we propose a lightweight and versatile hardware mitigation of SSB, named PseudoConflict. According to the evaluations through the RTL simulation and FPGA prototype experiment in terms of performance overhead and hardware resource utilization, the SSB attack can be effectively countered using the approach of PseudoConflict, even on a bare-metal processor with comparatively modest resources.
In order to effectively identify malicious activities in computer systems, Data Provenance-based analysis has been proposed to automatically correlate and visualize dependencies between events. However, a significant challenge known as “dependency explosion” arises when numerous benign activities are included in the generated graph, making it difficult to isolate attack-related activities. This paper proposes a novel method to address dependency explosion by extracting and removing patterns of frequently occurring benign activities using natural language processing and similarity-based analysis of log data. Unlike previous approaches that either exclude individual benign events or focus on extracting malicious activities, our method identifies benign activity patterns at an activity level without requiring frequent retraining. Experiments using the DARPA Transparent Computing Dataset demonstrate that approximately 6.8
Retrieval-augmented generation (RAG) is a technique that generates responses to questions by leveraging information retrieved from databases and large language models (LLM). However, with the advancements of LLMs and RAG, there are growing concerns about the potential for unintended responses resulting from malicious attacks, such as prompt injection attacks. To use LLMs and RAG techniques more safely, it is important to understand the vulnerabilities inherent in RAG by examining various attack methods. This paper proposes an untargeted adversarial input attack method specifically aimed at RAG. The proposed method involves adding adversarial strings to the prompts that are used as queries for RAG. By optimizing the adversarial strings to minimize the similarity between the prompt with adversarial strings and the relevant information contained in RAG’s database during information retrieval, we can effectively reduce the accuracy of the generated responses. Additionally, by applying a poisoning attack that injects sentences with adversarial strings into RAG’s database, we further decrease the accuracy of the responses. Through evaluation experiments, we confirmed that the proposed adversarial input attack and the poisoning attack were successful across multiple models and datasets.
Document summarization has long been a key task in natural language processing, with numerous methods proposed over the years. The advent of Transformer models has significantly improved performance for general-purpose summarization. However, specialized domains often require fine-tuning, which can be costly due to the need for curated training data. In this paper, we propose a document summarization framework that leverages named entity recognition (NER), focusing on cyber threat intelligence. We show that effective summarization is achievable using only publicly available data. Furthermore, we demonstrate that our proposed framework enables summarization incorporating generative AI such as GPT, and that it can generate effective summaries compared to simple zero-shot prompting summarization.
In wireless communication, adversarial jamming seeks to render messages incomprehensible for legitimate parties, friendly jamming aims to protect communication from illegitimate eavesdroppers. In both cases, it is crucial to understand if jammed signals may be recovered, either as a defensive or an offensive measure. The prevailing assumption is that effective signal recovery requires at least as many antennas as data sender and jammer, who use separate antennas, have in total. We challenge this convention by showing that the effects of jamming can be effectively neutralized using only one antenna when Binary Phase-Shift Keying modulation is used which is a mandatory modulation as per the IEEE 802.15.4 specification (the basis for Zigbee or WirelessHART among others). Our work builds on detailed simulation and practical experiments. In simulations, Bit Recovery Rates (BRRs) are mostly near 0.90; in over-the-air tests with a 15 dB stronger jamming signal, the average BRR is 0.70. Promising results of a follow-up experiment with a communication distance of 10 m underscore our method’s feasibility in the real-world. Our code and data are publicly available for transparency and to enable replicability.
