Information Security

Differential-linear cryptanalysis is a well-known cryptanalytic method combining differential and linear cryptanalysis. Since its introduction, it has become one of the most important tools for analyzing block ciphers. This paper focuses on differential-linear key-recovery attacks and presents a more efficient key-recovery algorithm by incorporating the partial-sum technique. This algorithm enables the key recovery attack to be divided into multiple steps, and the time complexity of a differential-linear key-recovery attack can be significantly reduced by carefully treating each step. Using this algorithm, we propose the first 19-round differential-linear key-recovery attacks on the message processing phase of GIFT-COFB and HyENA, which are currently the best-known attacks against these ciphers. Additionally, we extend the differential-linear attack on GIFT-64 to 19 rounds, surpassing the previous differential-linear attack by one round. We note that the attack results in this paper are far from threatening the security of GIFT-COFB, HyENA, and GIFT-64.

We propose a new physical-layer encryption scheme through pilot designs and MIMO techniques. Under formal reduction proofs and informal reliability analysis, we demonstrate that the decoding complexity for the legitimate user grows linearly with the number of antennas, whereas for the eavesdropper, decoding is computationally infeasible. In comparison to traditional wireless physical-layer security schemes, the proposed scheme remains secure even if the eavesdropper has unlimited computing power, infinite antennas, or knows the legitimate channel. Our scheme leverages the physical-layer wireless channel’s properties to achieve sophisticated network-layer encryption without a pre-shared key. Additionally, our algorithm involving lattices provides a new approach for secure group communication and can be used to enhance the security of the post-quantum cryptosystem in 6G. Experimental results show that the proposed scheme achieves a bit error rate of approximately 0.5 for the eavesdropper and nearly 0 for the legitimate receiver.

GCM with Secure Short Tag (GCM-SST) is a variant of the GCM authenticated encryption mode designed for improved security with short tags, and its standardization is ongoing in various organizations, including 3GPP and IETF. The original design specification was published with informal security claims only, and Inoue et al. then verified them with formal security proofs. They proved that the term regarding tag length t is \(\frac{v}{2^t}\) in GCM-SST (cf. \(\frac{v \ell }{2^t}\) in GCM), wherein v is the number of decryption queries and the maximum message block length \(\ell \). However, the proofs were given in the single-user (su) setting only, and its multi-user (mu) security remained an open research problem, which is a significant gap because GCM-SST’s specification document explicitly considers mu use cases and even recommends nonce randomization (NR) for improving mu-security. Addressing this issue, this paper proves mu-security of GCM-SST, verifying that the security with short tags stays intact under the mu setting. Moreover, by combining GCM-SST with NR and nonce-based key derivation (NKD), we show that those enhancement methods improve mu-security in the same level as those combined with GCM.

With the increasing deployment of encrypted protocols such as TLS, traditional deep packet inspection techniques have become ineffective, posing challenges to traffic classification. In this paper, we propose SimSeq, a robust TLS traffic classification method that relies solely on packet length sequences. To simulate real-world network conditions, we design two perturbation scenarios that emulate fast retransmission and timeout retransmission behaviors. Each scenario is configured with multiple packet loss rates, where smaller rates represent mild congestion and larger rates reflect more congested network conditions. We generate perturbation views of packet sequences using reliable transmission logic and leverage a contrastive learning framework to learn robust and discriminative representations. The encoder, composed of a BiLSTM and attention pooling module, is pretrained with a SimCLR-style contrastive loss and then finetuned with scenario-specific classification heads. Experimental results on the CESNET-TLS22 dataset show that SimSeq achieves strong and stable performance under both scenarios, with average F1-scores of 0.88 and 0.93, respectively.

Stateful Coverage-Based Greybox Fuzzing (SCGF) is a key technique for securing stateful network protocols. To efficiently process feedback and guide mutations, these fuzzers predominantly employ scheduling strategies based on simple short-term heuristics. However, the reliance on myopic heuristics, which fail to adopt a global, long-term optimization perspective, results in inefficient state-space exploration and a struggle to uncover vulnerabilities requiring deep and complex state transitions. To address this issue, we present EvoFuzz, an adaptive scheduling framework that applies Evolutionary Game Theory (EGT). EvoFuzz operates through two core modules: EvoState and EvoSeed. The EvoState module treats states as competing players to guide global exploration, while the EvoSeed module treats candidate seeds as competing players, selecting the most promising one within a target state. We implemented EvoFuzz on top of NSFuzz and evaluated it on five real-world protocols. The results indicate that compared to the state-of-the-art baselines, EvoFuzz increases the unique state sequence by up to 205.56%, increases code branch coverage by up to 6.87%, and increases the number of unique crashes by 31.96%.

In a previous study, Henrich et al. (ISC ’23) demonstrate how Transport Layer Security (TLS) handshake performance is affected not only by different Post Quantum Cryptography (PQC) Key Encapsulation Mechanisms (KEMs) and security levels, but also by varying physical network conditions. In particular, they show that prior to selecting a PQC scheme replacement for TLS, it is important to conduct an analysis of the anticipated network conditions for applications that require a high level of responsiveness. In this paper, we build upon the aforementioned work and complement the previous experiments to include digital signature PQC schemes and hybrid variants, as well as various compositions of certificate chains. Moreover, an analysis is conducted on the effects of deploying real physical servers and varying the underlying network stack configuration. Our results show that incorporating PQC signature schemes does not negatively impact the overall transmission time as substantially as poor network conditions. However, operating at high security levels frequently results in delays using PQC schemes. These findings are consistent across hybrid schemes as well. We conclude that migrating TLS to PQ-only or hybrid usage can generally be undertaken with a high degree of confidence. However, considering suboptimal network conditions or the use of higher security levels, a cautious transition is recommended. In such cases, the configuration of certificate chains or increasing the Transmission Control Protocol (TCP) Congestion Window (CW) might prove beneficial.

Attribute-based proxy re-encryption (AB-PRE) is a crucial variant of proxy re-encryption. It allows a proxy with a re-encryption key to transform a delegator’s ciphertext associated with an access policy into another ciphertext associated with a new access policy, enabling delegatees with matching attributes to decrypt the transformed ciphertext. However, a key limitation of AB-PRE is that the delegator cannot control which ciphertexts are transformed. As a result, the proxy, once given the re-encryption key, indiscriminately transforms all ciphertexts, effectively switching their underlying policies—an issue known as the all-or-nothing problem. It limits the system’s flexibility and practicality in real-world use cases.

In this paper, we address this by proposing a primitive of Conditional AB-PRE (CAB-PRE), which extends AB-PRE by incorporating conditional re-encryption. In CAB-PRE, the proxy can transform a ciphertext only if this ciphertext satisfies a specific condition set by the delegator in the re-encryption key. We formalize the adaptive security of CAB-PRE under the context of honest re-encryption attacks (HRA). We also give a concrete construction based on the learning with errors (LWE) assumption, which attains designated security in the standard model.

In this paper, we construct the first lattice-based threshold ring signature scheme with signature size scaling logarithmically in the size of the ring while supporting arbitrary thresholds. Our construction is also concretely efficient, achieving signature sizes of less than 150 kB for ring sizes up to \(N = 4096\) (with threshold size \(T=N/2\), say). This is substantially more compact than previous work.

Our approach is inspired by the recent work of Aardal et al. (CRYPTO 2024) on the compact aggregation of Falcon signatures, that uses the LaBRADOR lattice-based SNARKs to combine a collection of Falcon signatures into a single succinct argument of knowledge of those signatures. We proceed in a similar way to obtain compact threshold ring signatures from Falcon, but crucially require that the proof system be zero-knowledge in order to ensure the privacy of signers. Since LaBRADOR is not a zkSNARK, we associate it with a separate (non-succinct) lattice-based zero-knowledge proof system to achieve our desired properties.

Side-channel attacks are critical as they, despite the mathematical security of the algorithm, break the security assumption that private data stays hidden from the adversary. Developing secure hardware can be expensive, as multiple iterations of prototyping may be required to achieve a satisfactory level of security against side-channel attacks. Currently, the fairly new and open-source CPU-platform RISC-V is gaining traction by entering the Internet of Things (IoT)- and consumer market and also gains interest in security oriented projects such as OpenTitan. In case of security-critical applications, especially when the hardware is exposed to third party, the implementations of cryptographic algorithms must be secure against side-channel attacks. For the RISC-V platform currently only a small number of tools exist to assess the probing security. Further, we could identify a lack of simulation-based tooling to do so, with the ability to analyze larger implementations as e.g., full ciphers. To address this demand, we use PROLEAD_SW as a starting point and extend it to support the RISC-V platform. By analyzing micro-architectural leakage effects on the RISC-V platform we show that the CPU-independent leakage model used by PROLEAD_SW for the ARM architecture is suitable for the RISC-V platform. To verify the correctness of the new tooling, test-vectors are executed with the new tooling. In a final step, the performance of the new tooling is compared to the performance of the original version of PROLEAD_SW by analyzing two masked AES C implementations with both tools.

Conflict-based side-channel attacks allow attackers to monitor victims’ access patterns by asserting malicious cache conflicts. While cache randomization has emerged as a potential defense, existing solutions face critical limitations. CEASER-S and DT4+EV10 fail to fully prevent existing eviction set searching algorithms. MIRAGE suffers from intolerable area and power overheads. Chameleon’s relocation mechanism faces the problem of excessive power/energy consumption. To alleviate these limitations, we employ a dual-mapping randomized cache with global indirect replacement (GIR-Cache). A randomized direct-mapped look up table is designed to eliminate dual-index checking overhead by maintaining the active mapping state of each LLC address. Our approach effectively mitigates conflict-based side-channel attacks while incurs negligible runtime performance impact with moderate area and power overhead.

Online voting enables individuals to participate in elections remotely, offering greater efficiency and accessibility in both governmental and organizational settings. As this method gains popularity, ensuring the security of online voting systems becomes increasingly vital, as the systems supporting it must satisfy a demanding set of security requirements. Most research in this area emphasizes the design and verification of cryptographic protocols to protect voter integrity and system confidentiality. However, other vectors, such as network traffic analysis, remain relatively understudied, even though they may pose significant threats to voter privacy and the overall trustworthiness of the system.

In this paper, we examine how adversaries can exploit metadata from encrypted network traffic to uncover sensitive information during online voting. Our analysis reveals that, even without accessing the encrypted content, it is possible to infer critical voter actions, such as whether a person votes, the exact moment a ballot is submitted, and whether the ballot is valid or spoiled. We test these attacks with both rule-based techniques and machine learning methods. We evaluate our attacks on two widely used online voting platforms, one proprietary and one partially open source, achieving classification accuracy as high as 99.5%. These results expose a significant privacy vulnerability that threatens key properties of secure elections, including voter secrecy and protection against coercion or vote-buying. We explore mitigations to our attacks, demonstrating that countermeasures such as payload padding and timestamp equalization can substantially limit their effectiveness.

The rise of online social networks has heightened concerns over image privacy leakage. Although deep learning methods have been applied to privacy recognition, they face two key challenges: (1) a privacy gap between low-level visual features and high-level, context-aware human judgments, and (2) limited consideration of inter-entity context. To address these, we propose MSPP-Net, a Multi-Stage Privacy Perception Network inspired by human cognition. It decomposes privacy inference into three stages: entity perception to detect key objects, attribute perception to align visual features with semantic concepts via multimodal contrastive learning, and privacy perception to model inter-object context using graph attention networks. Experiments on our FineViP dataset show that MSPP-Net outperforms strong baselines, improving mAP by 3% and OR by 1.1%, validating the benefits of structured, cognitively motivated modeling for privacy recognition.

With the rapid development of pre-trained language models, data privacy has become a critical concern. Federated Parameter-Efficient Fine-Tuning has emerged as an effective solution, preserving privacy while controlling computational and communication costs. However, the joint participation of data owners in the training process makes it vulnerable to backdoor attacks, which drives our focus on backdoor attacks in the Federated Parameter-Efficient Fine-Tuning.

Experiments show that using efficient fine-tuning methods to freeze a large number of parameters does impact the success rate of backdoor attacks. Specifically, when parts near the output layer are frozen, the success rate of the backdoor attack significantly decreases, while the main task still converges normally. Based on these findings, we propose a new backdoor attack method: Frozen Layer Adversarial Sample-based Enhancement method. This method first generates adversarial examples that manipulate the output of frozen layers to target a specific class. Then, the trainable parameters are fine-tuned to generate these adversarial examples when backdoor data is input. Our experiments on GLUE text classification and CIFAR-10 image classification demonstrate that even when the server freezes parameters near the output layer, our method ensures a high success rate for backdoor attacks while maintaining stealth.

Photo Response Non-Uniformity (PRNU) noise serves as a sensor-level fingerprint in camera-based authentication and source attribution systems. Rather than degrading or suppressing PRNU patterns as in prior work, we introduce a novel transfer attack that injects PRNU noise from one device into images from another source or generated by AI. This enables manipulated images to falsely pass forensic source verification checks, posing a new class of threat to PRNU-based authentication. Our method achieves an average 85.5% compromise rate, validated using both a custom PRNU injection pipeline and the commercial forensic tool (Amped Authenticate). We further propose two mitigation techniques to detect such spoofing, revealing critical limitations in current image forensics pipelines.

Fuzzy Extractors (FEs) and Fuzzy Signatures (FSs) are promising primitives for template-protected biometric authentication, and lattice-based constructions of them are known. In this paper, to reveal lattices more suitable for FEs/FSs in terms of application to biometric authentication, we evaluate the accuracy of FEs/FSs for various lattices, along with the computation time of finding the closest lattice vector \(\textrm{CV}_L(\cdot )\) required in the authentication process when FEs/FSs are applied to biometric authentication. Specifically, we treat the integer lattice \(\mathbb {Z}^n\), a triangular lattice \(L_n^{(\textrm{tri})}\), and the direct product \(E_8^{n/8}\) of the Gosset lattice, which have been treated in conventional studies on FEs/FSs, and additionally the dual lattice \(L_n^{(\mathrm {d-tri})}\) of a triangular lattice and the checkerboard lattice \(D_n\). To evaluate the accuracy of FEs/FSs with these lattices, we give algorithms for computing the lattice norm for \(L_n^{(\mathrm {d-tri})}\), \(D_n\), and \(E_8^{n/8}\), where the lattice norm can be utilized for efficient accuracy evaluation and algorithms for \(\mathbb {Z}^n\) and \(L_n^{(\textrm{tri})}\) are known. Then, we evaluate the accuracy of FEs/FSs based on these lattices utilizing the lattice norm. Although \(L_n^{(\textrm{tri})}\) is often used for FEs and FSs conventionally, the evaluation results show that \(E_8^{n/8}\) achieves the highest accuracy of the evaluated lattices, and \(D_n\) achieves accuracy close to \(L_n^{(\textrm{tri})}\) with shorter computation time of \(\textrm{CV}_L(\cdot )\). Also, to obtain the lattice norm for \(L_n^{(\mathrm {d-tri})}\), we give a similarity transformation from a non-full-rank lattice to a full-rank one, which transforms the zero-sum root lattice \(A_n\) and its dual \(A_n^*\) to \(L_n^{(\textrm{tri})}\) and \(L_n^{(\mathrm {d-tri})}\), respectively. Using this transformation, we discuss a relation between \(L_n^{(\textrm{tri})}\), often used for FEs/FSs, and \(A_n\), a well-studied lattice in lattice theory.

The Internet of Things (IoT) devices are increasingly exploited as intermediaries for launching sophisticated cyberattacks. IoT honeypots have emerged as a proactive measure to lure attackers and provide early threat detection. However, existing honeypots exhibit significant limitations, including low interaction levels, vulnerability to fingerprinting, and constrained data collection capabilities. This paper introduces HoneySentry, a high-interaction IoT honeypot specifically designed to overcome these challenges and target advanced attackers adept at sophisticated honeypot fingerprinting and strategic selection of victim IoT devices. HoneySentry utilizes a custom-enhanced IoT firmware emulation framework to achieve high-fidelity emulation of diverse IoT devices and architectures. It incorporates advanced anti-fingerprinting techniques to evade detection, modifying commands frequently used by attackers during reconnaissance. Additionally, HoneySentry enhances its appeal to attackers by deploying a variety of meticulously crafted bait files and processes. To facilitate detailed analysis, HoneySentry captures comprehensive attack data, including both network traffic and host-level activities. Comparative evaluations against traditional honeypots and real-world deployments demonstrate that HoneySentry significantly outperforms existing solutions in fostering deep engagement with attackers, collecting extensive attack data, and enabling comprehensive threat analysis. During its two-month deployment (August 2024 to November 2024), HoneySentry captured over 200,000 requests and generated 61.3 GB of log data. Further analysis revealed variants of known malicious worms and viruses, as well as several intriguing attack behaviors, highlighting its capability to uncover diverse threats.

IoT malware is often created by modifying publicly available source code, resulting in numerous variants. Analyzing the functionality of these variants has become increasingly important. Function Call Sequence Graph (FCSG) have been proposed to represent internal function calls in binaries, offering a promising approach for functional analysis. However, the structure of FCSGs is highly sensitive to CPU architecture and compiler optimization, hindering cross-architecture analysis. In this paper, we propose a method for generating architecture-independent FCSGs by removing obstructive functions—such as initialization routines and architecture-specific functions—that are not called in the source code but appear in conventional FCSGs. Our method removes, on average, 97.4% of such functions across binaries compiled for Arm, i586, and MIPS architectures with different optimization levels. Furthermore, we show that the resulting FCSGs better reflect the similarity of the original source code, as measured by graph- and string-based similarity metrics. These results demonstrate the potential of our method to improve the robustness and consistency of IoT malware functional analysis across diverse architectures.

Research on Android malware has progressed rapidly, yet the task of distinguishing malicious from benign applications continues to test the limits of automated analysis. Earlier work, dominated by static signature matching, frequently struggles when novel or obfuscated samples appear. Contemporary graph–based pipelines alleviate some of these shortcomings by modelling control and data dependencies, but their reliance on pairwise relations often blurs higher–order interactions that experienced adversaries nurture when crafting evasive variants. These observations motivate a return to first principles: we require representations that faithfully encode behaviours without incurring prohibitive overhead. In this study we revisit the problem through the lens of hypergraph representation learning. Treating an application as a hypergraph allows one to encode joint behaviours—such as the co-invocation of critical API calls inside a single execution context—that cannot be decomposed into simple edges without information loss. Building on this representation, we introduce HGANN-Mal, a Hypergraph Attention Neural Network that adaptively emphasises semantically salient hyperedges while softening the influence of spurious ones. The model derives its signals from static analysis to extract structural and semantic features. Importantly, on the Drebin dataset, HGANN-Mal achieves a Macro-F1 score of 97.8% and an accuracy of 98.3% in binary malware detection, significantly outperforming graph-based and static hypergraph methods. Our findings validate that the proposed attention-based hypergraph model provides a more exhaustive and precise solution for detecting sophisticated Android malware.

Interpreting the massive volume of security alerts is a significant challenge in Security Operations Centres (SOCs). Effective contextualisation is important, enabling quick distinction between genuine threats and benign activity to prioritise what needs further analysis. This paper proposes a graph-based approach to enhance alert contextualisation in a SOC by aggregating alerts into graph-based alert groups, where nodes represent alerts and edges denote relationships within defined time-windows. By grouping related alerts, we enable analysis at a higher abstraction level, capturing attack steps more effectively than individual alerts. Furthermore, to show that our format is well suited for downstream machine learning methods, we employ Graph Matching Networks (GMNs) to correlate incoming alert groups with historical incidents, providing analysts with additional insights.

Modern Endpoint Detection and Reaction (EDR) systems must remain reliable even when attackers acquire high privileges within the monitored operating system. Ideally, such systems should be protected from compromise, while maintaining deep visibility into the internal behavior of the target system. To meet this challenge, we aim to isolate the monitored system in a virtual machine (VM) and move the EDR logic into the hypervisor. We present HyperSec, a domain-specific language that allows the VM to safely delegate detection logic to the hypervisor by transmitting specialized monitoring programs. This design bridges the semantic gap by enabling the hypervisor to understand OS-level structures, while constraining accepted programs to protect the hypervisor against potential vulnerabilities. We evaluated our system against kernel rootkits that hide processes or elevate privileges and showed that HyperSec enables reliable detection of such threats. Our results indicate that the performance overhead remains acceptable, paving the way for broader adoption of hypervisor-based EDR with custom in-VM insight.

Bootkits threaten the very foundation of system security. By exploiting vulnerabilities in firmware and bootloaders, these attacks gain persistent, stealthy control at the earliest stages of boot. Despite widespread adoption of UEFI Secure Boot and TPM-based measurements, limited visibility and complexity during early boot allow bootkits to evade detection. Ongoing discovery of critical firmware flaws enables privilege escalation and circumvention of core protections. Existing static verification methods cannot detect runtime modifications during early boot, highlighting the need for runtime-aware integrity monitoring. In this paper, we present BootMarker, a runtime integrity monitoring framework built on a dual-layered architecture that combines Driver Execution Environment (DXE) and System Management Mode (SMM) instrumentation. BootMarker dynamically enforces control-flow integrity in the bootloader and performs cryptographic validation of firmware components in real time. It detects and mitigates bootkit attacks as they occur during early execution. Our evaluation shows that BootMarker reliably identifies diverse bootkit behaviors while imposing minimal performance overhead, making it practical for real-world deployment and significantly enhancing boot-time security.

In the context of software supply chain security and IoT device firmware analysis, binary vulnerability detection faces dual challenges of detection efficiency and coverage due to scarce annotated binary data. Although the open-source ecosystem has accumulated vast amounts of source-level vulnerability data, direct migration to binary vulnerability detection inevitably encounters a semantic gap caused by cross-modal representation differences such as compiler optimizations and symbol stripping. To address data scarcity in binary vulnerability detection and bridge the semantic gap in cross-modal matching with source code, this paper proposes a hierarchical semantic fusion framework for binary-source alignment. Through heterogeneous modal semantic bridging and hierarchical attention mechanisms, our approach significantly enhances cross-modal matching precision and scalability between binary and source code, achieving 94.3% accuracy. Furthermore, we introduce a vulnerability detection task-driven transfer framework that maps source-level vulnerability patterns to binary code feature space via cross-modal alignment. Leveraging dimensional expansion within the model’s knowledge space enables exponential scaling of usable data for binary vulnerability detection, thereby transcending data scarcity constraints. We collected 400 CVEs from 8 real-world vulnerable projects, achieving 80.3% detection accuracy. This research establishes an effective technical pathway for expanding usable data resources in automated binary vulnerability detection.

In this work, we introduce CryptNyx, a Password Hardening (PH) framework that enhances the security of stored password records through collaboration between the authentication server and an external server, also known as rater. PH mitigates offline password brute force attacks on stolen databases by involving the rater in the password verification process, enabling it to impose limits on password decryption attempts. However, this means that the remote server can track user login requests, raising concerns about potential compromises to user privacy. Consequently, achieving effective rate-limiting while preserving user anonymity has remained an unresolved challenge.

CryptNyx ensures anonymity without sacrificing rate-limiting. Essentially, the user pseudonym, which allows the rater to track login requests, can be refreshed any number of times in a controlled but unlinkable manner, offering complete anonymity while still mitigating offline guessing attempts. Furthermore, CryptNyx allows for password-hardened encryption capabilities, which enable users to securely encrypt sensitive data using their strengthened password records. Additional features include an “Opt-out” protocol that facilitates client withdrawal, and an “Anonymous Opt-in” protocol designed for efficient batch registration.

Experimental results demonstrate the effectiveness and practicality of our approach, highlighting the balance between user privacy, security, and system functionality.

In dynamic group signature schemes (GSS), forward security ensures that newly joined members cannot generate valid signatures for past time periods. Additionally, non-frameability prevents even privileged entities, such as the group manager or key issuer, from falsely attributing signatures to honest users. Most GSS either lack non-frameability or face significant efficiency challenges when updating signing keys to ensure forward security. In this paper, we introduce a forward-secure dynamic group signature scheme that guarantees non-frameability. We also present an alternative scheme that, while lacking non-frameability, offers higher efficiency compared to existing schemes with comparable security. For both protocols, we propose efficient revocation mechanisms that allow an authority to revoke users without requiring re-registering existing users. Additionally, we propose a technique that enables the verification process of both protocols to be performed in batches. We prove the security of our schemes, ensuring the standard dynamic GSS security notions; anonymity, traceability and non-frameability (second scheme). Experimental results demonstrate that our schemes are competitive in both computational and communication efficiency when compared to existing literature.

Continuous authentication helps mitigate the risk of session hijacking, insider attack, and privilege abuse. Applying zero trust principles, this paper proposes a family of four formally specified access control models to account for the use of continuous authentication to monitor user access patterns in user-facing software applications, each model providing increasingly expressive user modeling capabilities. We name these models Zero Trust Continuous Authentication (ZTCA). Deploying a ZTCA model requires the authoring of policies. To ease the challenge of developing ZTCA policies, we studied the problem of automatically generating ZTCA policies from declarative usability and security requirements. We devised a novel SAT encoding for the automated policy formulation problem, so that policy formulation can be performed by state-of-the-art SAT solvers. Empirical experiments demonstrate that our novel encoding approach runs significantly faster than a competing encoding approach previously published in the literature.

With the rapid development of blockchain technology, vulnerabilities in smart contracts have become a major threat to asset security. Traditional rule-based detection methods, although interpretable, often suffer from high false positive rates and limited scalability. Despite recent progress, deep learning methods are often limited to unimodal approaches and lack the capability for fine-grained analysis. Our analysis of 14 common vulnerabilities revealed that function-level granularity strikes the optimal balance between detection accuracy and efficiency. Building on this observation, we propose the Jakiro method, which improves detection accuracy by integrating the semantic information from control flow graphs (CFGs) and source code using cross-modal contrastive learning. Experiments conducted on a dataset of more than 38,000 real-world contracts demonstrate that Jakiro surpasses the majority of the 10 baseline methods across three tasks: reentry, integer overflow, and transaction order dependency, achieving average improvements of 6.49%, 2.66%, and 4.62% in precision, recall, and F1 score, respectively.

This paper presents BlockLens, a supervised, trace-level framework for detecting malicious Ethereum transactions using large language models (LLMs). Unlike prior approaches limited to static features or storage-level abstractions, BlockLens processes complete execution traces, capturing opcode sequences, memory information, gas usage, and call structures to accurately represent the runtime behavior of each transaction. This framework harnesses the exceptional reasoning capabilities of LLMs for long input sequences and is fine-tuned on transaction data. We design a tokenization strategy aligned with Ethereum Virtual Machine (EVM) semantics, mapping execution traces into interpretable tokens. Each transaction captures its complete execution trace through simulated execution and is then sliced into overlapping chunks using a sliding window, allowing for long-range context modeling within memory constraints. During inference, the model outputs both a binary decision and a probability score indicating the likelihood of malicious behavior. We implement the framework based on LLaMA 3.2-1B backbone and fine-tune the model using Low-Rank Adaptation (LoRA). We evaluate it on a curated dataset containing both real-world attacks and normal DeFi transactions. BlockLens outperforms representative baselines, achieving higher F1 scores and recall at top-k thresholds than representative baselines. Additionally, BlockLens offers interpretable chunk-level outputs by localizing suspicious trace segments that enhance explainability, facilitating rapid forensic analysis and actionable decision-making in security-critical environments.