Data valuation provides a principled framework for quantifying the contribution of data to model training. It plays a crucial role in trustworthy machine learning (ML) by supporting data curation, enhancing interpretability, and enabling fair incentive mechanisms in data markets. Shapley value is a popular method for data valuation, but accurate estimation remains computationally expensive, particularly at the dataset level. In this paper, we introduce Ensemble Shapley, an efficient framework tailored for dataset-level valuation on the sharded structure. To reduce the computational costs, we propose a two-phase estimation method that apportions the intensive contribution computation costs across disjoint data shards and strategically reuses the computation results, achieving efficient contribution evaluation through the ensemble of shard models. However, weak shard models trained on noisy data may degrade ensemble models' performance. To solve this, we introduce a behavior-driven guided sampling method that pairs noisy datasets with benign ones, ensuring reliable contribution estimates despite the noise. We also derive an advantageous lower bound for the number of evaluation iterations that balances efficiency and accuracy by the number of shards. Experimental results show Ensemble Shapley has superior efficiency over existing methods while maintaining comparable accuracy across various ML tasks, and demonstrates strong scalability and integration potential.

As blockchain technology is increasingly being applied to various aspects of actual production, higher requirements are being placed on its stability and correctness. As the core logical structure of the blockchain system, the consensus protocol bears the pressure of coordinating, exchanging, and processing information among all parties in the system, and it also requires further improvement. Currently, the improvement and development of the consensus protocol have reached a level that can effectively cope with the pressures and requirements of the current situation. However, with the development of computing technology, the computing power of all parties will continue to improve, and the threat to the consensus protocol will also increase simultaneously. Traditional simulation verification methods may no longer be able to meet existing requirements. More scientific verification methods should be employed when designing protocols to ensure accuracy and reliability. Additionally, a unified induction of the current verification level for each protocol is necessary to facilitate a more comprehensive analysis of future research ideas. This paper begins with an overview of consensus protocols, gathering and summarizing extensive research on the use of formal methods to verify protocol formulas. It analyzes the common properties of these verification protocols, along with their definitions and formal expressions. The paper then examines the methods and tools commonly employed for formal verification, discussing the characteristics of each approach, current research trends, and the challenges that remain to be addressed. Through the analysis of specific case studies, we outline the historical context and current achievements in this area, providing valuable references for future research.

Gradient leakage attacks pose a significant privacy threat in distributed learning-as-a-service APIs. Existing literature on gradient leakage defense relies on gradient perturbation for preventing privacy leakage. However, determining where and how much to perturb the gradient offers different capabilities for preventing gradient leakage. This paper presents GradCloak, a principled approach to guiding gradient perturbation with theoretical robustness bounds in federated learning as a service, aiming to find the minimum required noise for simultaneously achieving privacy protection, competitive accuracy, and preventing gradient leakage attacks. The paper is organized into three major components. First, we formulate the gradient leakage threats and their adverse effect. We categorize the attack into two broad types: leakage during local training and leakage before global aggregation. Second, we investigate different gradient perturbation approaches. We analyze and compare these gradient perturbation methods, which are performed at the federated server, with those performed at the participating client(s). Third, we introduce three robustness properties of robust perturbation against gradient leakage threats, formulated by the anonymization bound for training data robustness, the perturbation bound for gradient robustness, and the distribution robustness bound for perturbed gradients. We conduct extensive evaluations on eight benchmark datasets to demonstrate that specific settings of gradient perturbation exist that best balance privacy, accuracy, and leakage prevention.

The proliferation of cloud-based healthcare services has created a growing demand for secure and fine-grained sharing of Personal Health Records (PHRs). To support layered access control, Hierarchical Attribute-Based Keyword Search (HABKS) schemes have been developed. However, existing designs suffer from inherent collusion vulnerabilities that is users with access to different subtrees can jointly recover higher-level secrets, breaking access boundaries. To address this problem, we pro pose a collusion-resistant hierarchical multi-keyword searchable encryption scheme for privacy-preserving PHR sharing. Different from the previous HABKS, our scheme effectively prevents the propagation dependence of data encryption keys at nodes and user collusion attacks. This benefits from the root node key decoupling mechanism we designed, which enables each policy subtree to generate keys with embedded randomness. Additionally, we design an identity-bound trapdoor and a dual verification mechanism that tightly bind user attributes and keyword tokens, ensuring verifiable and fine-grained access control. Our scheme supports expressive multi-keyword queries with constant-size trapdoors, and offloads most pairing operations to the cloud, requiring only one pairing locally. Security proofs and experimental results confirm that our scheme achieves strong collusion resistance, efficient keyword retrieval, and practical suitability for secure hierarchical PHR systems.

In the era of Industry 4.0, distributed storage systems face significant challenges in terms of reliability, data security, privacy, and maintenance. While traditional solutions like HDFS rely on centralized nodes that create single points of failure, existing blockchain-based alternatives often lack efficient mechanisms to dynamically evaluate node reliability in untrusted environments, leading to potential data unavailability.This study introduces a secure, blockchain-based distributed storage system using the Trusted Adaptive File Allocation (TAFA) Algorithm for efficient file distribution. The system uses smart contracts to automatically incentivize and monitor nodes, rewarding or penalizing them to boost long-term availability. Extensive simulation experiments demonstrate that this system achieves an optimal Load Balance Factor (LBF) of nearly 1.00 and maintains high storage reliability. It is noteworthy that, even in environments with up to 90% malicious nodes performing tampering or denial-of-service attacks, the file availability of this system remains at least 30%, which is up to 5.7 times higher than that of other distributed storage algorithms.

How much influence can a coordinated coalition exert in a multiwinner Top-k election under a positional scoring rule? We study the maximum displacement problem: with coalition size m, how many of the current top-k winners can be forced out? We show coalition power decomposes into two independent prefix-majorization constraints, capturing how much the coalition can (i) boost outsiders and (ii) suppress weak winners. For arbitrary scoring rules these prefix inequalities are tight, efficiently checkable necessary conditions (exact in the continuous relaxation). For common-step arithmetic-progression (AP) score ladders, including Borda, truncated Borda, k-approval/k-veto, plurality, and multi-level rules such as 3–2–1, we prove a Majorization–Lattice Theorem: feasible aggregate score vectors are exactly the integer points satisfying the Block–HLP prefix-sum capacity constraints plus a single global congruence condition modulo the step size g. For Borda (g=1) the congruence vanishes, yielding a pure prefix-majorization test. This characterization yields an O(k'log k') exact feasibility oracle for displacing k' winners, and an O(k(log k)^2log(mx)) algorithm (via dual-envelope binary search) for computing the maximum achievable displacement k^∗. Experiments on Mallows profiles and PrefLib elections confirm exact cutoffs, diminishing returns, and substantial gains over baseline heuristics; for g>1 they also demonstrate the predicted congruence effect, where prefix-only tests produce false positives. The oracle scales to extreme instances, processing 10^9 candidates in under 28 seconds (memory permitting).

Manual formal analysis of cryptographic schemes is labor-intensive and requires substantial expertise. While model-checking tools (e.g., Scyther and Tamarin) and computational-security tools (e.g., CryptoVerif and EasyCrypt) improve the automation of security proofs, they still rely on experts to abstract schemes and write tool-specific formal descriptions. Large language models (LLMs) are a promising alternative, but their effectiveness in this domain remains unexplored due to the absence of standardized evaluation methodologies. To fill this gap, we introduce CrypFormBench (C.F.B for short), a comprehensive benchmark jointly covering symbolic and computational security to evaluate five core LLM capabilities: interpretation, generation, completion, transformation, and correction. It comprises 700 instances spanning 677 schemes, 7 mainstream formal verifier languages, and 160 security properties. The evaluation of 9 state-of-the-art LLMs reveals that most of them perform well on interpretation and completion, given their code-awareness advantages, but struggle with generation, transformation, and correction. Overall, their performance remains limited, with Claude-3.5 achieving the highest score at 48.7 out of 100. We further provide practical guidance, e.g., few-shot prompting, Pass@K sampling, and lightweight fine-tuning, to mitigate the executability bottleneck and improve tool-usable outputs. Taken together, our benchmark and analyses offer a grounded view of current progress and concrete directions toward reliable LLM-assisted formal cryptographic analysis.

Network traffic classification is crucial for network management, security monitoring, and ensuring quality of service. However, classifying traffic from Decentralized Applications (DApps) is challenging due to their similar encryption configurations on the same blockchain platform, making their traffic patterns indistinguishable. Traditional classification methods, such as Deep Packet Inspection, struggle with encrypted traffic. Similarly, existing machine learning methods, like Random Forest (RF), fail to effectively capture the unique traffic patterns of DApps. To address these challenges, this paper proposes a novel multimodal classification approach that combines Graph Neural Networks (GNN) and Bidirectional Encoder Representations from Transformers (BERT) to classify DApp traffic. This method leverages GNN and BERT to extract and analyze interaction patterns and payload information from network traffic, using self-attention mechanisms to fuse these features, providing a comprehensive feature representation. We collected traffic data from 100 commonly used DApps on the Ethereum platform to create a high-quality dataset. Experimental results on this dataset demonstrate that our model achieves 90.96

As a variant of the signcryption scheme, threshold unsigncryption operates among multiple receiver members through a (t, n) threshold process during the unsigncryption phase to improve security and robustness, and has been widely deployed in various applications such as cloud computing, blockchain, and electronic auctions. However, existing schemes only consider security under a limited static corruption model, where the adversary selects which parties to corrupt before the protocol begins. As a result, these schemes lack the flexibility to address more complex attack forms posed by real-world adversaries. To solve this issue, it is worth investigating threshold unsigncryption schemes with adaptive security. In this paper, we propose the first provably secure (t, n) threshold unsigncryption scheme in the standard model against adaptive corruption adversaries who can corrupt parties at any time during protocol execution. The main contributions include: First, we provide the formal syntactic definition and security model of threshold unsigncryption schemes under the adaptive corruption model. Second, we employ the dual system technique to formally prove that the proposed scheme can achieve adaptive security in the standard model. Notably, our scheme is equipped with valuable properties such as non-interactivity, robustness, public verifiability, and splittable unsigncryption, significantly enhancing usability and functional completeness. Besides, we implement the proposed scheme on two types of elliptic curve groups with different security levels and evaluate its practical performance under different thresholds. Experiments demonstrate that on a composite order group with 80-bit security, a single receiver only takes 16.61 s to perform the threshold unsigncryption phase when the number of receivers n is 20 and the threshold t reaches up to 18. As a pioneering adaptively secure threshold unsigncryption scheme, our work not only establishes a benchmark for future proposals but also broadens its application in scenarios requiring high-security levels and flexible key management, such as electronic voting.

Federated learning (FL) has emerged as a promising paradigm for privacy-preserving distributed machine learning, yet it remains highly vulnerable to model poisoning attacks. Malicious clients can submit manipulated updates to degrade model performance or implant hidden backdoors. While robust aggregation and attack detection have been extensively studied, efficiently recovering the global model after poisoning remains challenging. Retraining from scratch incurs prohibitive costs, whereas rollback-based approaches fail to eliminate incremental contamination. To overcome these limitations, we propose FedReconPP, a lightweight recovery framework. We design an importance-driven compressed storage mechanism to significantly reduce overhead. We introduce a time-decayed contamination tracking mechanism with rigorous theoretical foundation that captures and mitigates incremental poisoning and develop a quality-aware privacy-preserving recovery strategy with formal differential privacy guarantees to strengthen robustness against inference attacks while maintaining model utility. Extensive experiments on multiple datasets, demonstrate that FedReconPP achieves reliable recovery with substantially lower resource consumption, offering practicality and strong protection for federated learning systems.

Federated Learning (FL) is a transformative paradigm for training global machine learning models using decentralized datasets hosted by edge or client devices, without requiring centralized data aggregation. This makes FL particularly valuable for privacy-sensitive applications deployed in edge-cloud environments. However, FL is vulnerable to Byzantine attacks, where malicious clients provide falsified local model updates to compromise the performance of the global model. Such vulnerabilities are especially problematic in heterogeneous and resource-constrained edge-cloud systems, where ensuring trust across distributed clients is a significant challenge. This paper introduces a novel Byzantine-Robust FL method designed to address these challenges in edge-cloud computing scenarios. Our method leverages a central server equipped with a small clean dataset as an initial root of trust to evaluate the trustworthiness of client updates. Unlike prior approaches, our method iteratively builds a dynamic set of trusted clients, gradually incorporating their updates into the training process to refine the global model. This dynamic trust mechanism reduces reliance on the quality of the initial clean dataset, ensuring robustness even in the presence of a significant number of malicious clients. Extensive experiments on real-world datasets demonstrate the effectiveness of our approach in achieving high model accuracy and maintaining robustness against Byzantine attacks. Our method is particularly well-suited for edge-cloud environments, addressing critical challenges such as resource constraints, distributed learning management, and the reliability of collaborative AI systems.

Local Differential Privacy (LDP) has been widely recognized as a powerful tool for providing a strong theoretical guarantee of data privacy to data contributors against an untrusted data collector. Under a typical LDP scheme, each data contributor independently randomly perturbs their data before submitting them to the data collector, which in turn infers valuable statistics about the original data from received perturbed data. Common to existing LDP mechanisms is an inherent trade-off between the level of privacy protection and data utility in the sense that strong data privacy often comes at the cost of reduced data utility. Frequency estimation based on Randomized Response (RR) is a fundamental building block of many LDP mechanisms. In this paper, we propose a novel Joint Randomized Response (JRR) mechanism based on correlated data perturbations to achieve locally differentially private frequency estimation. JRR divides data contributors into disjoint groups of two members and lets those in the same group jointly perturb their binary data to improve frequency-estimation accuracy and achieve the same level of data privacy by hiding the group membership information in contrast to the classical RR mechanism. Theoretical analysis and detailed simulation studies using both real and synthetic datasets show that JRR achieves the same level of data privacy as the classical RR mechanism while improving the frequency-estimation accuracy in the overwhelming majority of the cases by up to two orders of magnitude.

Quantum Federated Learning (QFL) has emerged as a promising research direction by combining the strengths of quantum computing and federated learning. However, existing QFL solutions have consistently failed to simultaneously improve client training efficiency and ensure communication security. In this paper, we present a novel Multi-qubit Broadcast-based QFL framework (MB-QFL) to address the efficiency and security challenges of existing approaches. The framework employs a novel multi-qubit broadcast protocol and a quantum average method to secure the information transmission process. The multi-qubit broadcast protocol overcomes the limitations of existing protocols by allowing the transmission of an arbitrary S-qubit state from one sender to multiple (Q) receivers, whereas earlier protocols were restricted to broadcast one or two qubit state to recipients. Additionally, we propose an averaging method for quantum states, which exploits the probabilistic cloning technique to achieve aggregation in MB-QFL. The security analysis demonstrates that MB-QFL can effectively protect against inference attacks from malicious clients, as well as eavesdropping and intercept-and-resend attacks during communication. The algorithm complexity of MB-QFL is significantly lower than existing QFLs. Besides, the experimental results indicate that MB-QFL achieves higher classification accuracy than other QFLs.

Anomaly-based network intrusion detection systems (NIDSs) are essential for ensuring cybersecurity. However, the security communities realize some limitations when they put most existing proposals into practice. The challenges are mainly concerned with fine-grained unknown attack detection and ever-changing legitimate traffic adaptation. To tackle these problem, we present three key design norms. The core idea is to construct a model to split the data distribution hyperplane and leverage the concept of isolation, as well as advance the incremental model update. We utilize the isolation tree as the backbone to design our model, named, to echo back three norms. By analyzing the popular dataset of network intrusion traces, we show that significantly outperforms the state-of-the-art methods. Further, we perform an initial deployment of by working with the Internet Service Provider (ISP) to detect distributed denial of service (DDoS) attacks. With real-world tests and manual analysis, we demonstrate the effectiveness of to identify previously-unseen attacks in a fine-grained manner.

Distributed Denial of Service (DDoS) defense is a profound research problem. In recent years, adversaries tend to complicate their attack strategies by crafting vast DDoS variants. On the one hand, this trend exacerbates both extremes of classification granularity (i.e., binary and attack level) in existing machine learning methods. On the other hand, massive attack categories make the filter rule table bulky, as well as cause problems of slow reaction presented in the recent state-of-the-art DDoS mitigation system. Therefore, we propose the concept of a DDoS family to reconcile/cope with these issues. The specific technical roadmap includes traffic pattern characterization, attack fingerprint production, and cross-executed family partition by community detection. Through extensive evaluations, we demonstrate the benefits of the proposal in terms of portraying similarities, guiding model classification/unknown attack detection, optimizing defense strategies, and speeding filtering reactions. For instance, our results show that using only one rule can defend 15 types of attacks due to their homogeneous behavioral representation. Particularly, we find the interesting observation that counting the backward packet is more efficient and robust against some attacks (e.g., Tor's Hammer Attack), which is very different from previous solutions.

Recent studies show that deep neural networks are extremely vulnerable, especially for adversarial examples of image classification models. However, the current defense technologies exhibit a series of limitations in terms of the adaptability of different attacks, the trade-off between clean-instance accuracy and robust one, as well as efficiency for train time overhead. To tackle these problems, we present a novel component, named redundant fully connected layer, which can be combined with existing model backbones in a pluggable manner. Specifically, we design a tailor-made loss function for it that leverages cosine similarity to maximize the difference and diversity of multiple fully connected parts. We conduct extensive experiments against 12 representative attacks (white-box and black-box), based on the popular dataset. The empirical evaluations show that our scheme realizes significant outcomes against various attacks with negligible additional training overhead, while hardly bringing collateral damage for clean-instance accuracy.

The continuous emergence of malware has threatened to the Android platform and user privacy. With the evolution of the Android system and malware, it is challenging to design a method that can accurately identify the categories of sophisticated malware, including known and unknown families, as well as their obfuscated variants, given that they may be newly emerging and lack available detection knowledge. Although some methods try to use anomaly detection and zero-shot technology to identify unseen applications, they are limited to binary classification or lack the robustness, stability, universality, and interpretability in multi-class identification. To this end, we first propose a generic meta-features mining algorithm, which can discover the potential relationships between samples belonging to the same category. Then we present metaNet, a novel method leveraging meta-features to identify sophisticated Android malware. Specifically, metaNet is mainly powered by four components: (i) mExtractor is a feature collector to obtain the static and dynamic features. (ii) mProcessor is taking unique meta-features of each category from extracted features. (iii) mLearner is a machine learning suite that leverages features and meta-features to design and train a classifier called HSU-Net. (iv) mEnforcer is a flexible deployer that identifies categories of malware families in the real world. We implement a prototype of metaNet with 15K lines of Python code and compare it with state-of-the-art (SOTA) methods. The results show that it can not only achieve superior performance in terms of known families (99.52% of accuracy) and unknown families (99.31% of accuracy trained with 80% known families) for binary classification, but also perform well in multi-class identification, i.e., 99.05% and 93.45% of accuracy for known and unknown families, respectively. Furthermore, we deploy and evaluate metaNet in the real world. It can identify applications over an acceptable time and memory overheads, i.e., average of 11.8s and 56MB per sample with a size of 8MB. Also, the few-shot detection and feature perturbation experiments reflect its robustness and stability benefiting from meta-features. Finally, we collect the traffic of 112 decentralized applications (DApps) belonging to 16 categories, such as finance and health, and evaluated metaNet in DApp identification. The results illustrate its applicability across various tasks. That is, it can accurately classify 94.6% and 81.36% of DApp flows in all-known and 80%-known DApp scenarios, respectively, outperforming the SOTA methods.

This study proposes a network security information transmission encryption model based on blockchain, aiming to solve the security and privacy protection problems of information transmission in modern networks. The model combines the decentralized characteristics of blockchain and multi-party secure computing (MPC) technology to ensure confidentiality and integrity during data transmission. First, a transparent and tamper-proof encrypted information transmission link is established through blockchain technology, which effectively prevents the risk of data tampering and forgery. Secondly, the encryption algorithm based on multi-party secure computing further enhances the data privacy protection between the participants, ensuring that no single node can obtain all the information during the data transmission process. Through simulation experiments, the model’s anti-attack ability under various network attacks and its performance in different transmission environments are verified. The simulation results show that the model has significant advantages in the security and stability of data encryption transmission, especially in the face of man-in-the-middle attacks and denial of service attacks. It has a good defense effect. Experimental data show that its transmission encryption efficiency is about 15% higher than that of existing methods.

Ethereum, one of the most prominent and widely deployed blockchain systems, is undergoing a significant upgrade that adopts sharding for capacity expansion and secure multi-party computation (MPC) to enable distributed validator technology (DVT). However, it faces a data availability problem, where an adversary can cheat honest-but-lazy validators to propagate invalid blocks, thereby exposing Ethereum Sharding to vulnerabilities. To address this issue, proof of custody (PoC) was proposed. Regrettably, no practically effective MPC protocol is designed to securely and distributively compute PoC. In this paper, we present a concrete and efficient MPC protocol that enables the secure computation of PoC in the dishonest-majority malicious setting. First, we construct an efficient conversion protocol to convert secret sharings in two different prime fields using the doubly authenticated bits (daBits) technique. Second, an efficient MPC protocol is designed to compute scalar multiplication over an elliptic-curve group without requiring MPC to compute branching programs. Furthermore, we employ affine coordinates to compute group operations over an elliptic curve, leading to significant performance boosts compared to other coordinate systems. Finally, a concrete end-to-end implementation of the protocol is built, and its performance is evaluated. When the operations of a validator are collaboratively performed by seven parties, the online (resp., total) running time to generate one proof of custody is 0.02 s (resp., 559.6 s) in LAN and 0.39 s (resp., 7904.2 s) in WAN. The results demonstrate the practicality of the proposed protocol for Ethereum Sharding over a long period of time, such as 73 days.

The frequent security incidents of contracts indicate a pressing need to ensure contract security from deployment to running stages, but the state-of-the-art (SOTA) analysis methods cannot work well for three requirements. (i) Identify contract defective code snippets, while generating exploit call sequences to help developers fix them. (ii) Monitor abnormal call behaviors, especially for multiple continuous transactions. (iii) Validate numerous unexploitable detection results automatically because manual verification is labor-intensive. To tackle these problems, we propose SymX, a symbolic execution-based security analysis art accounting for contract development and running stages. The experiment results demonstrate that it can accurately identify 90.22% of contracts and 98.04% of call transactions, as well as validate misreports as intended, which is superior to SOTAs, thereby protecting contracts better during the contract lifecycle. Currently, SymX is available at https://github.com/Secbrain/SymX.