In real-world cyberattacks, adversaries frequently exploit a combination of vulnerabilities, bugs, and misconfigurations to compromise systems. To systematically analyze the root causes behind these issues, the Common Weakness Enumeration (CWE) framework provides a standardized taxonomy of software weaknesses. While vulnerability databases are central to cataloging known issues, many security-relevant descriptions first appear in informal sources such as blog posts, CTI reports, and social media. Although these sources predominantly offer broader cybersecurity insights, they occasionally yield details that may indicate underlying weaknesses not captured in formal databases. We propose a two-step approach to extract these security-related descriptions from unstructured threat intelligence and automatically map them to their corresponding CWE categories. First, a binary classifier detects sentences resembling CVE descriptions, identifying information relevant to security teams. Then, we apply a self-supervised learning model to predict the most appropriate CWE, enabling structured analysis even in the absence of formal vulnerability tracking. As no ground truth exists for this task, we conduct expert-driven validation. Our results show strong performance, with an F1-score of 98.17
Mapping Common Vulnerabilities and Exposures (CVEs) to the related Common Weakness Enumerations (CWE) is crucial in cybersecurity because this link allows categorizing vulnerabilities, prioritizing remediation efforts, and improving mitigation strategies. However, this process often requires manual effort to understand and connect a CVE to its CWE, which is undesirable due to its time-consuming nature. Existing automated approaches typically rely on vectorizing CVE descriptions using embedding techniques followed by machine learning classifiers. However, little attention has been given to evaluating whether CVE descriptions are the most effective starting point for automated classification, as these descriptions often contain irrelevant details that do not contribute to CVE-to-CWE mapping. Our research investigates to what extent we can automatically extract key information from CVE descriptions required to perform such mapping and evaluates how this technique improves existing methods for identifying related weaknesses using this extracted information. To this end, we present an approach that automates extracting key terms of CVEs through Large Language Models and evaluate the effect of focusing on various parts of the CVE description. We show that our key term extraction technique improves the F1-score of transformer-based classification of CVEs into CWEs up to 8.88
The NTRU problem with multiple keys arises when a user generates multiple distinct NTRU public keys h_i = g_i/f q by keeping its private key f fixed, and an adversary tries to recover the user’s private key f by collecting multiple instances of public keys h_i . At DCC 2023, Kim et al. introduced a polynomial-time algorithm to address this problem. Subsequently, at ICICS 2024, Song et al. identified a weakness in Kim et al.’s approach and proposed their own method. There are various versions of the NTRU algorithm, resulting in different ways of selecting the parameter f . For instance, in the NIST PQC standardization process, the NTRUEncrypt in Round 1 and NTRU-HPS in Round 3, chooses the private key f in the form f = 1 p to improve efficiency. However, previous studies focus on the scenario where the private key f is a ternary polynomial with a fixed Hamming weight. In this paper, we propose a polynomial time attack against NTRU with multiple keys, assuming that f=1+p*F , where F is a ternary polynomial of degree at most N-1 with a fixed Hamming weight and p=3 . Given the fixed Hamming weight of g_i , we can derive the initial modular equations from different h_i . By employing linearization techniques and exploiting the relationships between the coefficients of the equations, we can establish a system of linear modular equations with ⌊N/2⌋ variables that possesses a unique solution. Using this solution, we can effectively recover the private key f . We conducted experiments using the parameters of NTRU scheme in NIST PQC Competition and achieved successful results with an overwhelming probability. As a by-product, we can naturally extend the multi-key attack scenario to the NTRU-based proxy re-encryption scheme NTRUReEncrypt proposed by Nu ñ ez et al. at AsiaCCS 2015, and successfully recover the user’s private key.
Communication between satellites and mission control centers is currently protected via the Space Data Link Security (SDLS) Protocol. This protocol uses exclusively symmetric cryptography, which makes it resilient against quantum computers, but prevents key updates from achieving post-compromise security and limits operational efficiency, especially in future federated environments. To resolve these issues we designed a key-update/establishment protocol based on PQNoise that leverages KEMs to achieve confidentiality and authenticity. We prove the resulting scheme secure in a slightly modified version of the Bellare–Rogaway-model. To keep the SDLS protocol untouched, the resulting protocol is not used as the handshake phase of a larger data-transmission protocol, but actually run independently, permitting to start from shared keys.
Card-based cryptography performs a secure computation using a deck of playing cards through a series of physical actions, such as shuffling a sequence of cards. In 1989, the first card-based protocol called the five-card trick was proposed to compute the two-input AND function. The five-card trick needs a practical shuffling action called a random cut only once, which shifts a sequence of cards by a random offset. Subsequent research aimed to extend the five-card trick to compute any Boolean function and to reduce the number of cards required to compute it, but Mizuki, Kumamoto, and Sone in 2012 showed that the five-card trick itself can be done with four cards. This Mizuki–Kumamoto–Sone protocol uses the minimal number of cards required for computing two-input Boolean functions (as long as we encode an input bit with two cards); however, it needs a practical step of shuffling action twice, and the question of minimizing the number of steps, particularly the number of shuffling, remains an open problem. In this study, we negatively answer this problem; we prove that any four-card AND protocol cannot be realized using only a single practical shuffle. This implies that the Mizuki–Kumamoto–Sone protocol utilizes the minimal number of practical shuffles. For this, we enumerate all possible practical shuffles and prove that applying any one of them only once either cannot compute the AND function or leak information about the input.
Designated verifier signatures (DVS) enable a signer to create a signature that only a specific verifier can authenticate, thereby ensuring both message authenticity and signer privacy. Strong designated verifier signatures (SDVS) further enhance the security of the signer’s identity. In particular, they guarantee that it is computationally infeasible for any third party, even those possessing the secret keys of the signer or the designated verifier, to determine the public key under which the signature was produced. In this paper, we introduce a novel construction of a strong designated verifier signature scheme based on isogeny assumptions. Our scheme leverages the computational difficulty of finding isogenies between supersingular elliptic curves, a problem believed to be resistant to quantum attacks. Our construction achieves strong off-the-record security, unforgeability, and privacy of the signer’s identity in the random oracle model. Furthermore, we employ optimized techniques to substantially reduce key sizes without compromising signature compactness. The resulting public keys are only 16 kB, and the secret keys are 128 bits, while a signature achieving 128-bit classical security is 862 bytes, which is smaller than those in lattice-based or code-based SDVS proposals.
A fundamental component of a Cold Boot Attack involves a reconstruction algorithm for recovering corrupted encryption keys. This paper introduces a novel heuristic designed to significantly augment the efficiency of such algorithms. Our work illustrates the heuristic’s capacity to streamline key reconstruction across various encryption key types and integrates it into an AES Encryption key reconstruction algorithm. Subsequently, we perform an empirical assessment to gauge the performance of this algorithm compared to current state-of-the-art methods.
A recent study by Yamashita and Yasunaga (GameSec 2023) presented a constant-round deterministic broadcast protocol secure against detection-averse adversaries—those who prefer to attack without being detected. In this work, we revisit their protocol and observe that it remains secure even against a broader class of adversaries, not necessarily detection-averse. We formalize its detection mechanism as local detectability and construct broadcast protocols with local detectability that address two weaknesses of the original protocol: (1) it only guarantees weak validity, and (2) it may cause false detections. Our first protocol achieves round complexity four against rational adversaries and t+4 against malicious adversaries, where the adversary corrupts at most t parties. Our second protocol achieves the optimal round complexity of t+1 for malicious adversaries, while the round complexity is four against detection-averse adversaries.
Recently, a previously unseen supply chain attack due to a backdoor in XZ Utils has been identified by A. Freund. This particular attack leverages highly sophisticated attack techniques, starting with social engineering attacks against the open source community up to implanting a backdoor in obfuscated binary blobs. This malware is designed to empower the attacker(s) to remotely run commands on vulnerable servers utilizing SSH evading authentication. A reverse dependency analysis revealed that the affected library is used by almost 30,000 packages in Debian and Ubuntu—the same order of magnitude as the GNU standard C library (glibc/libc6), a dependency for roughly 50,000 packages on these systems. This fact highlights the severity of such supply chain attacks and raises concerns about further backdoored packages still undetected in the wild. This paper identifies the critical attack path for successful implantation of such a backdoor and abstracts general key takeaways for future detection and mitigation of similar attacks. We also present SketchyCrawler, an open-source tool prototype designed to illustrate crawling repositories to reveal ‘sketchy’ signs of potential backdoor implantation attempts.
Gaston is a cryptographic permutation introduced by El Hirch et al. at Crypto 2023. It uses the same number of bitwise operations as Ascon-p, the core permutation of the NIST-selected standard for lightweight authenticated encryption Ascon. Despite this similarity in complexity, Gaston provides stronger resistance to differential and linear cryptanalysis, making it a promising option for lightweight cryptographic applications. In this work, we evaluate the hardware performance of Gaston and Ascon-p, and conduct power side-channel analysis on the initialization phase of Ascon when instantiated with each permutation on an FPGA. We propose a majority voting technique that leverages the structure of Ascon-p, where the diffusion layer follows the non-linear layer. This method improves the key recovery success rate for Ascon instantiated with Ascon-p, even with fewer traces than in previous studies. To explore the role of layer ordering, we also consider a modified version of Gaston, named Gaston-R, in which the non-linear and diffusion layers are swapped. While this modification does not affect theoretical security, it may influence both the hardware cost and the effectiveness of the specific side-channel attack applied in this study. Our hardware analysis shows a small overhead for Gaston-R, and the side-channel results indicate noticeably reduced key recovery success compared to the original Gaston. To our knowledge, this is the first work to investigate how layer ordering impacts the complexity of side-channel attacks of this kind. Our findings suggest that this aspect may deserve further exploration in the design of side-channel-aware cryptographic permutations.
The rapid evolution of enterprise network architectures from traditional monolithic systems to cloud-native platforms, such as Microsoft Azure, driven by its accessibility and cost-efficiency has introduced significant complexity and expanded the potential attack surface. As the boundaries of network perimeters blur within distributed cloud infrastructures, security solutions like lateral movement (LM) detection face significant challenges in keeping pace with evolving threats. Moreover, the scarcity of attack data hinders the development and testing of effective detection algorithms. To address this gap, we adapted the Hopper attack synthesis framework to model interactions among internal users, applications, and resources within cloud platforms. Our adaptation generates synthetic datasets that simulate various attack scenarios for Azure login data. These synthetic data serve as the foundation for our research, enabling the systematic evaluation of multiple detection approaches without the need for sensitive production data.
Splitting up sensitive data into multiple shares – termed masking – has proven an effective countermeasure against various types of Side-Channel Analysis (SCA) on cryptographic implementations. However, in software this approach leads to dramatic performance overheads for non-linear operations. This problem can be addressed with hardware accelerators particularly designed to process masked data securely. In this context, Gao et al. [24] presented a RISC-V Instruction Set Extension (ISE) with masked Boolean and arithmetic instructions to accelerate masked software implementations of block ciphers. In this work, we demonstrate how this ISE can be applied to and extended for Post-Quantum Cryptography (PQC) components, forming a crypto-agile solution. We provide masked implementations based on three different ISE constellations for multiple highly relevant components, including Cumulative Distribution Table (CDT) sampling and polynomial rotation. With the masked instructions, we measure speedups of up to a factor 4.2 compared to sophisticated bitsliced implementations and even up to more than two orders of magnitude for non-bitsliced implementations. We assert the first-order security of our new instructions with a practical evaluation.
Multi-party computation (MPC) has become increasingly practical in the last two decades, solving privacy and security issues in various domains, such as healthcare, finance, and machine learning. One big caveat is that MPC sometimes lacks usability since the knowledge barrier for regular users can be high. Users have to deal with, e.g., various CLI tools, private networks, and sometimes even must install many dependencies, which can be hardware-dependent. A solution to improve the usability of MPC is to build browser-based MPC engines where each party runs within a browser window. Two examples of such an MPC web engine are JIFF and the web variant of MPyC. Both support an honest majority with passive corruptions. : Our work brings one of the most performant and versatile general-purpose MPC engines, MP-SPDZ, to the web. MP-SPDZ supports ≥ 40 MPC protocols with different security models, enabling many security models on the web. To port MP-SPDZ to the web, we use Emscripten to compile MP-SPDZ’s C++ BackEnd to WebAssembly and upgrade the party communication for the browser (WebRTC or WebSockets). We call the new MPC web engine webSPDZ. As with the native versions of the mentioned MPC web engines, MPyC-Web and JIFF, webSPDZ outperforms them in our end-to-end experiments. We believe that webSPDZ brings forth many interesting and practically relevant use cases. Thus, webSPDZ pushes the boundaries of MPC: enabling MPC for a broader audience by making it more usable.
Ateniese et al. (CRYPTO 2019/JoC 2021) introduced a cryptographic primitive which they call matchmaking encryption (ME), and Identity-based ME (IB-ME) is its identity-based variant. IB-ME supports an equality matching where a sender (encryptor) indicates a receiver’s (decryptor’s) identity ( ) in addition to their own ID ( σ ), and a receiver indicates a sender’s identity ( ) in addition to the own identity ( ρ ). A ciphertext is decrypted if (σ ,ρ )=(,) . In this paper, we pay attention to the search condition of public key authenticated encryption with keyword search (PAEKS) (Huang-Li, Information Sciences 2017) is reminiscent of the equality matching. We introduce a public key variant of ME which we call PK-ME, and propose a generic construction of PK-ME from PAEKS. As a conceptual contribution, our work lies in revealing a connection between ME and public key searchable encryption, which were independently researched so far. Due to the generic construction of PAEKS (Li-Boyen, IACR CiC 2024), we can instantiate the proposed generic construction from pairings or lattices. Finally, we discuss the advantage/disadvantage of PK-ME compared to IB-ME.
Ciphertext-Policy Attribute-Based Encryption (CP-ABE) is a powerful cryptographic mechanism that enables fine-grained access control, making it widely used for secure data management and sharing in cloud environments. However, no existing CP-ABE scheme provides an efficient and comprehensive solution for verifying both key generation and encryption correctness, leaving errors or malicious behavior undetectable. Such vulnerabilities may result in unauthorized access and compromise the reliability and security of the system. In this work, we propose the first Fully Verifiable CP-ABE scheme, introducing mechanisms to verify the correctness of both key generation and encryption outputs. Our approach leverages pairing-based cryptography to allow users to confirm that their private keys correspond to their assigned attributes while enabling public verification that ciphertexts accurately reflect the intended access policies. By introducing verification mechanisms for both the user’s private key and ciphertext, our scheme not only ensures that CP-ABE functions as intended, but also significantly enhances system security, accountability, and reliability. Moreover, our construction preserves the efficiency, expressiveness, and security guarantees of the underlying CP-ABE scheme, making it well-suited for real-world applications that require both correctness and efficiency.
Binary code similarity detection (BCSD) is fundamental to various applications, including firmware vulnerability search, code clone detection, and malware code segment identification. Many deep learning models have been proposed to learn code semantics from graph structures. However, most existing methods struggle to capture long-range dependencies in binary code represented by tree structures. To address this issue, this paper proposes Lanstree, a cross-architecture binary code similarity detection framework based on a bidirectional Tree-Transformer. Lanstree performs bottom-up and top-down semantic learning on abstract syntax trees (ASTs), enabling each node to fully aggregate features from its parent and child nodes. It employs a global attention mechanism to obtain a comprehensive representation of the AST and adopts a Siamese architecture to enhance binary code similarity detection performance. We have implemented a prototype of Lanstree, and experimental results demonstrate that Lanstree outperforms previous state-of-the-art BCSD methods.
Vulnerabilities in source code can lead to a wide range of issues, such as privacy and security breaches, system crashes, data leaks, and unpredictable application behavior. These problems typically arise from improper handling of inputs, memory, or access control mechanisms. To address these challenges, we propose a novel Attention-based Code Summarization (ACS) approach for multi-label vulnerability detection in source code. Unlike traditional methods that are limited to binary or multi-class classification, our approach is designed to identify multiple types of vulnerabilities within a single code snippet—an essential capability, as real-world software often contains more than one vulnerability. Our method introduces a specialized, security-focused attention mechanism that utilizes vulnerability-specific queries to generate targeted representations for each vulnerability type. By extending pre-trained code models with a hierarchical summarization framework, we enable the creation of specialized representations tailored to different vulnerability categories. This design ensures a balance between general code understanding and precise vulnerability detection. Extensive evaluations on a multi-label vulnerability dataset show that our model outperforms existing baselines, especially for buffer-related vulnerabilities (e.g., CWE-119, CWE-120) and other complex security issues. Through comprehensive ablation studies, we validate the individual contributions of each architectural component, demonstrating the synergistic effects of our vulnerability-specific attention, security-focused summarization, and adaptive gating mechanisms. Furthermore, the attention visualization component of our model provides explainable outputs, helping security teams identify and understand the root causes of detected vulnerabilities. By enabling efficient and simultaneous detection of multiple vulnerability types, our approach significantly reduces the effort and time required to analyze and secure software systems.
Malware presents significant challenges to cybersecurity by exploiting system vulnerabilities, compromising sensitive data, disrupting operations, and rapidly evolving to evade traditional detection methods. Many types of malware, such as wipers, are challenging to detect due to their low computational footprint. This low activity level often causes existing detection methods, including those based on hardware performance counters (HPCs), to struggle. This paper investigates the impact of ransomware and wipers on various HPCs. Our study shows that wiper demands minimal computational activity due to its stealthy behaviour, which mostly affects only a small subset of HPCs, while most remain largely unaffected. Leveraging this insight, we propose improvements to supervised and unsupervised malware detection methods. For supervised detection, we introduce an ensemble-based algorithm to identify the most sensitive features (i.e., HPCs), leading to a 98
SAND-128 is an AND-RX-based Feistel cipher proposed by Chen et al. (Designs, Codes and Cryptography 2022). This paper evaluates the security of the lightweight block cipher SAND-128 against impossible differential attacks. We use the bit-level Constraint Programming (CP) model proposed by Hadipour et al., which enables an automatic search for impossible differential distinguishers without requiring fixed input/output differences. Applying this model, we identified 2^14× 7 = 114,688 distinguishers over 14 rounds. Using one of them, and following the complexity framework established by Boura et al., we show a 21-round key-recovery attack on SAND-128 with data, time, and memory complexities of 2^126 , 2^126.33 , and 2^122 respectively.
Accurate anomaly detection is essential for protecting heterogeneous multi-environment (M-En) networks, where traditional enterprise traffic coexists with Internet of Things (IoT) flows. Existing machine learning-based anomaly detection approaches rely on extensive labeled data and frequent retraining, and therefore adapt poorly to unseen attack patterns and few-data settings. In this work, a Retrieval-Augmented Generation (RAG) pipeline is introduced in which Large Language Models (LLMs) are guided to classify malicious traffic under few-shot settings. Structured packet-level and statistical features are first rendered into natural-language prompts; dense sentence embeddings are then produced and indexed with FAISS (Facebook AI Similarity Search, an approximate-nearest-neighbour vector index) to construct a vector-based knowledge base. At inference time, the embedding of an unseen packet is used to retrieve a specified number of semantically similar, labeled examples, which are supplied as in-context examples to the LLM. This workflow enables learning over heterogeneous traffic without model fine-tuning, improves generalization, and yields human-readable explanations of each decision. When paired with MPNet embeddings and a context of 200 retrieved examples, the 4B-parameter Gemma3:4b model reaches 1.0 accuracy, while the 7B Mistral model achieves 0.98. Inference takes 23.51 s with Gemma3:4b and 66.7 s with Mistral for 50 samples, both running on a single NVIDIA GeForce RTX 3090 GPU.
