With the widespread use of the Internet of Things (IoT) and mobile devices, binary programs are widely used in human life. From smart homes and car systems to industrial control systems, underlying firmware and native executable files form the core of these critical infrastructures. In traditional vulnerability discovery, researchers need to manually analyze function relationships and function content. With the rapid development of Artificial Intelligence (AI), large language models (LLMs) are utilized for analysis, replacing manual analysis. However, sending fully decompiled functions to LLMs consumes a large number of tokens, resulting in high costs and slow responses, particularly for large binaries. Many existing methods treat all functions equally, without focusing on those more likely to be vulnerable. In this paper, we proposed a token-efficient binary vulnerability prioritization framework. Our framework allows the LLM to analyze function names to identify potentially risky functions, and then enables the LLM to examine only these selected functions in detail. What’s more, we introduce a metric called token-detection efficiency (TDE) to better demonstrate the efficiency of our proposed framework. The experiment shows that our framework achieves higher vulnerability detection accuracy with fewer tokens.
A common approach to protecting data confidentiality again-st potential adversaries is to encrypt the databases. Nevertheless, it is infeasible to perform meaningful computations on a ciphertext. Trusted Execution Environments (TEEs) offer a promising alternative by enabling computations on plaintext within a secure enclave, thereby supporting rich and complex queries. However, existing works either fall short in providing strong oblivious guarantees, fail to support queries such as non-equi joins, or suffer from significant efficiency bottlenecks. In this paper, we first address the oblivious order counting and encoding problem in a parallelized manner, enabling the counting and encoding of elements that satisfy a given order within an unordered sequence. Building on this, we propose parallelizable non-equi join algorithms for both single condition and multiple conditions. The experimental results show that our scheme completes a non-equi-join with a single condition, producing 16,000,000 output records in 26.4 s, and a multiple conditions non-equi-join producing 26,366,724 output records in 19.4 s, achieving up to 4.7 × and 2.04 × speedups over the state-of-the-art scheme, respectively.
The CKKS scheme enables approximate arithmetic over encrypted vectors, supporting SIMD computation and controlled ciphertext rotations without revealing plaintexts. Prior encrypted fuzzy-query systems for English text typically map each character to a small, contiguous code and evaluate equality or similarity with low-degree polynomials. This design does not transfer to Chinese: the Unicode space is large and discontinuous, and after normalization for CKKS distinct characters can become numerically proximate, leading to collisions under approximation noise and undermining both equality tests and fuzzy comparisons. We address these limitations with two encoding strategies Digit Decomposition Embedding (DDE) and Bit Slicing (BS) that represent each Chinese character by several low-magnitude integer components, preserving inter-character separability at low multiplicative depth and remaining compatible with non-bootstrapping CKKS. We further introduce a slot-masking and sliding-window mechanism that leverages structured slot alignment and bounded rotations to realize fuzzy matching and substring retrieval. Together, these components provide a structured Chinese encoding framework for multi-field encrypted fuzzy query that maintains numeric distinguishability, controls depth growth, and exploits CKKS batch parallelism for efficient end-to-end execution.
Recently, growing public concern over data privacy has led to the emergence of the ‘right to be forgotten’ under the European Union’s General Data Protection Regulation (GDPR), which allows individuals to request the deletion of personal data. In response, researchers have developed methods to selectively remove or mitigate the influence of specific data on machine learning models, a process known as machine unlearning (MUL). However, while most research on MUL has focused exclusively on maintaining model accuracy after unlearning, the robustness of these models has been largely overlooked, especially under adversarial attacks. In this study, we examine whether unlearning degrades robustness in the visual classification task by performing adversarial attacks on various unlearned methods and compare performance against a retrained-from-scratch baseline, including evaluations using a standard corruption dataset. Our comprehensive evaluation across seven unlearning methods reveals consistent degradation in robustness under perturbation, supporting our hypothesis that machine unlearning degrades adversarial robustness compared to retraining from scratch, and highlighting the need for methods that preserve robustness without compromising the effectiveness and efficiency.
The proliferation of open-source software (OSS) has made software supply chains prime targets for attacks like Package Confusion, where adversaries publish malicious packages with names deceptively similar to legitimate ones. Existing detection methods often rely on simple lexical similarity or passive analysis of known package pairs, struggle with high false positive rates (FPR), fail to proactively identify emerging threats, and are vulnerable to adversarial evasion. To overcome these limitations, we introduce AgentGuard, a novel framework for proactive, single-input package confusion detection. AgentGuard employs a multi-agent architecture that autonomously discovers potential confusion targets using fine-tuned word embedding model to hybird semantic search and subsequently evaluates the risk via a machine learning model incorporating multi-dimensional feature groups to enhance robustness. This design enables scalable, real-time monitoring across diverse software ecosystems. We evaluate AgentGuard on the challenging ConfuDB and NeupaneDB datasets. Our results demonstrate that AgentGuard significantly outperforms state-of-the-art baselines, improving accuracy by 10
Recommender systems rely on user preferences and item attributes to provide personalized services, yet they inherently face a tension between model accuracy and data privacy. Vertical Federated Learning (VFL) allows multiple organizations to jointly train models without sharing raw data, typically using Private Set Intersection (PSI) to align overlapping users. However, conventional PSI supports only exact identifier matching and fails to accommodate the feature-based conditional filtering frequently required in modern recommendation pipelines. To overcome this limitation, we propose P ^2 FR-VFL, a VFL framework that integrates Privacy-Preserving Feature-Retrieval PSI (P ^2 FR-PSI). This enhanced PSI mechanism enables participants to privately align and selectively filter user identifiers according to hidden predicate conditions, without disclosing any sensitive feature information. This design enables P ^2 FR-VFL to provide a more expressive and privacy-preserving preprocessing stage for federated recommendation tasks. Building on this capability, P ^2 FR-VFL also offers a scalable and privacy-enhanced solution for cross-domain recommender systems, effectively reconciling model utility with strong user privacy guarantees in federated environments. Experiments on real-world datasets show that P ^2 FR-VFL achieves predictive accuracy comparable to plaintext VFL training, while incurring only 20% additional communication overhead relative to state-of-the-art PSI schemes, all while supporting flexible alignment capabilities.
With the rapid development and widespread adoption of the Internet of Things, cloud computing, and artificial intelligence, data security has become a critical concern. Homomorphic encryption, as an effective solution to privacy protection, enables computations to be performed directly on encrypted data without decryption. It has thus become a powerful cryptographic tool and is widely applied to secure aggregation in federated learning. However, as federated learning continues to evolve, the security of model transmission has emerged as a pressing challenge, and verifying the integrity and authenticity of the transmitted models remains an open problem. In this paper, we propose a verifiable and privacy-preserving federated learning (VPFL) framework, which simultaneously ensures privacy preservation and model verifiability while also supporting regulatory auditing. Based on a VPFL model trained on the MNIST dataset, our approach achieves a prediction accuracy of 90.67
Federated Learning (FL) has become a key paradigm for privacy-preserving collaborative machine learning, especially in sensitive fields such as medical imaging. However, large-scale, cross-institutional FL deployments still struggle with fine-grained access control, as most existing frameworks rely on static security policies that cannot adapt to participant churn or evolving regulatory requirements. To overcome these scalability and flexibility limitations, this paper presents a dynamic access control framework for FL based on Multi-Authority Attribute-Based Encryption (MA-ABE). We design a multi-attribute management architecture that enforces granular authorization over both local updates and global model parameters. Central to the framework is a policy-token mechanism that allows dynamic policy updates based on contextual changes without incurring high-cost re-encryption or disrupting training. We further introduce a context-aware key generation protocol to ensure secure, seamless policy transitions. Formal analysis shows that the scheme achieves CPA semantic security and resists collusion attacks. Experiments demonstrate that our protocol significantly reduces computational, communication, and storage overheads compared to state-of-the-art approaches, offering a practical and robust foundation for secure collaborative learning in dynamic environments.
Side-channel analysis evaluates cryptographic device security, but single channel methods can overlook combined leakage threats. Multi-channel fusion attacks exploit leakage more effectively. In this paper, we propose a decision-level fusion analysis method based on deep learning and Dempster-Shafer evidence theory, specifically tailored for side-channel analysis of public-key algorithms. To evaluate the reliability of sample classification probability distributions, we introduce a metric called the average separability index. Compared to data-level fusion and feature-level fusion, our method yields higher accuracy and confidence for cryptographic operations. In the side-channel analysis of ECC, RSA, and module-lattice-based key encapsulation mechanisms, key recovery accuracy is significantly improved, while the number of traces used is notably reduced. This approach achieves more than 98
Watermark-based detection shows strong potential for identifying machine-generated content, but applying it to code generation is still hard due to low entropy and strict structure, which limit flexibility and cause a trade-off between quality and detectability. Existing methods like SWEET use fixed entropy thresholds. They cannot adjust to different levels of token uncertainty across coding contexts. To overcome this limitation, we propose Entropy-Aware Watermarking (EAW), a framework that dynamically modulates watermark embedding strength based on local entropy signals. Specifically, EAW amplifies the greenlist bias in high-entropy regions, such as variable names or comments, while reducing interference in low-entropy, syntax-critical parts to maintain the functional correctness of the code. This adaptive mechanism enables EAW to maintain code naturalness without sacrificing detection robustness. Extensive experiments across multiple benchmarks demonstrate that EAW consistently outperforms prior baselines in both watermark detectability and code quality retention, establishing a better trade-off between transparency and reliability in code watermarking.
Passwords remain the primary authentication method in computer systems and are expected to prevail for the foreseeable future. Previous works on password security have mainly focused on the passwords created by English-speaking users. As a result, the proposed password guessing methods are confined to the password composed by English speakers. However, language structure plays a significant role in password creation, analysis that overlooks the language-specific inherent features may yield biased results. To address this limitation, we perform a large-scale empirical study using 64.5 million passwords from seven real-world webs. By using the cosine similarity measure, we find significant differences in character distribution between Chinese and English websites. In addition, we explore the use of Chinese Pinyin chunks and English words in both Chinese and English passwords, analyzing their impact on cracking rates. Finally, we propose a large language model (LLM) based password-guessing method to analyze the vulnerability of Chinese web passwords. Our experimental results demonstrate that incorporating linguistic features into the model improves the cracking rate by 3.97 10^7 guesses, our approach achieves a cracking rate of 37.21
Multimodal classification in medical imaging faces two key challenges in cross-modal feature fusion: significant semantic gaps between modalities often cause loss of critical pathological information during fusion, and existing diagnostic models generally process cases individually without leveraging population-level diagnostic knowledge. To address these issues, we propose the Dual-Feature Interaction Graph Neural Network (DFI-GNN). Our approach integrates feature interaction fusion, collective knowledge mining, and collaborative optimization. Specifically, we first introduce a dual-feature interaction fusion module that preserves original discriminative features through differentiable regularization constraints and residual feature backpropagation. Second, we construct patient association graphs based on multimodal feature similarity to mine population-level diagnostic knowledge via graph attention networks. Finally, a multimodal-graph collaborative learning strategy jointly optimizes multimodal loss, graph relational loss, and fusion regularization loss, enhancing single-modal discriminability, collective knowledge relevance, and cross-modal fusion robustness. Extensive experiments on multiple benchmarks show that DFI-GNN achieves superior performance in multimodal feature learning, offering a more reliable solution for medical image diagnosis.
The rapid proliferation of artificial intelligence (AI) systems has led to the generation of massive amounts of valuable data, which can be leveraged to enhance analytical capabilities and support intelligent decision-making. Since the quality of such decision-making is fundamentally determined by the AI models trained on these data, it is crucial to perform periodic audits to verify and maintain data integrity. However, traditional data auditing solutions encounter significant challenges due to the reluctance of different domains to share data. To this end, we propose a Blockchain-based Cross-Domain Data Auditing (BCDDA) scheme designed for trustworthy AI systems. BCDDA employs a Chameleon Hash-based mechanism to efficiently construct audit metadata, thereby reducing the computational overhead on domain servers. Furthermore, it adopts a parallel data auditing architecture, in which audit tags are recorded on the main blockchain, while AI data are redundantly stored across multiple domain servers to ensure recoverability even if a shard is compromised. In addition, BCDDA supports dynamic operations on AI data, such as insertion, deletion, and modification, without regenerating audit tags, thereby enhancing system flexibility. Experimental evaluations conducted with four shard consensus nodes demonstrate that the proposed BCDDA scheme achieves high efficiency in audit tag generation and shard consensus processes. The results confirm that BCDDA provides a secure, reliable, and scalable solution for cross-domain AI data auditing in decentralized environments.
Medical Internet of Things (MIoT), also known as smart healthcare or Medical 4.0, typically refers to intelligent devices such as medical equipment, wearable sensors, and remote monitoring instruments connected to healthcare applications. These devices are used to collect, transmit, exchange, and analyze health data, thereby improving the quality, efficiency, and safety of medical services. However, these medical data are often vulnerable to security threats during transmission. To safeguard secure communication and privacy preservation in MIoT, several certificateless aggregate signature (CLAS) schemes have been introduced. Nevertheless, most of these schemes depend on a centralized key generation center (KGC) and remain susceptible to coalition attacks, KGC compromise attacks, and distributed denial-of-service (DDoS) attacks. In this paper, we design a blockchain-based certificateless anonymous aggregate signature (CLAAS) scheme for MIOT and demonstrate that it can achieve complete decentralization of entities, thereby resisting KGC compromised attacks. This blockchain-based CLAAS scheme has strong robustness, which can resist DDoS attacks and coalition attacks launched by several medical sensor nodes. The proposed scheme can safeguard provable security in the random oracle model and no longer rely on secure channels to obtain the partial private key. Furthermore, performance evaluation reveals that the decentralized CLAAS scheme outperforms existing schemes in terms of both communication and computational overhead, highlighting its strong practical applicability in MIoT systems.
Efficient recycling of transparent PET bottles necessitates a multi-model AI system capable of concurrently performing material classification and component detection. However, executing multiple models sequentially on resource-constrained edge devices results in cumulative inference times. In this paper, we address this challenge by implementing and evaluating a GStreamer-based parallel inference pipeline on a Raspberry Pi CM 5 equipped with a Hailo-8 AI accelerator. The implemented parallel architecture replicates input frames into two independent branches, enabling simultaneous execution of a PET/CAN classification model and a cap/ring/label detection model. Experimental results demonstrate that the parallel pipeline reduced the average frame processing time by 11.3
Distributed learning is widely regarded as an effective solution for edge computing in the current data-intensive era. By eliminating heavy data transfers inherent to centralized machine learning, it enables participants to train models locally without aggregating raw data on a cloud server. However, data at distributed clients is often missing or tampered with, which leads to inaccurate gradients during training and opens the door to falsified or manipulated updates. To address the issue that heterogeneous data quality among clients in distributed learning may degrade the performance of the global model, this paper proposes an admission and continuous verification mechanism for federated learning based on Proof of Data Possession (PDP). In this mechanism, the proof of data possession serves as a prerequisite for client participation. Each client must demonstrate the integrity and authenticity of its designated dataset, thereby filtering out clients with incorrect or unreliable data. Furthermore, a continuous random sampling verification strategy is innovatively introduced during the federated training process, where clients that repeatedly fail the proof are eliminated. In practical deployment, considering the differences in computational capabilities among devices, they are categorized into high-capacity and low-capacity devices. A hierarchical PDP mechanism is designed accordingly: high-capacity devices perform integrity verification using BLS signatures, while low-capacity devices adopt sampling-based verification. The scheme further integrates aggregated proofs to reduce time overhead and improve overall efficiency.
Secure comparison protocols are a cornerstone of privacy-preserving computation, with critical applications in e-voting, federated learning, and financial analytics. However, the efficiency of classical protocols like DGK is often hampered by high computation and communication costs. This paper introduces an efficient two-party secure comparison protocol, termed the PSC (Prefix-based Secure Comparison) protocol, based on a variant of the ElGamal homomorphic cryptosystem. The core innovation of PSC is to process each prefix bit string of the input as a single encrypted unit, bypassing the need for costly bit-wise equality checks found in traditional approaches. The protocol determines the comparison result by constructing and analyzing an obfuscated comparison vector for zero-value occurrences. If we denote the bit string length by l , this method reduces the computational complexity by approximately 𝒪(l) and minimizes interaction rounds. Experimental evaluations confirm that PSC achieves a significantly lower average running time compared to the DGK protocol and its variants. The results demonstrate that reorganizing the encrypted data model enables substantial efficiency gains in secure comparison without sacrificing security, offering a practical solution for large-scale privacy-preserving tasks.
BLS signature scheme is widely employed in blockchain systems due to its compact signature size and non-interactive aggregation capability. However, the standard BLS threshold signature faces critical bottlenecks. First, the verification process needs computationally expensive bilinear pairing operations. Second, the distributed key generation requires multiple communication rounds. This work introduces a novel fast-verifiable threshold BLS signature scheme. We design a non-interactive key generation protocol that integrates program obfuscation with pseudorandom functions. It enables participants to deterministically derive threshold keys for dynamic groups through a single registration, which eliminates interactive rounds. Second, we develop an efficient verification algorithm by embedding Chaum-Pedersen proofs, which bypasses the conventional bilinear pairing verification. It can shifts computational overhead to more efficient elliptic curve operations. Third, we provide formal security proofs in the random oracle model under standard cryptographic assumptions, including the co-CDH hardness assumption. It demonstrates existential unforgeability against chosen message attacks. Experimental results demonstrate that our scheme achieves over 50
When detecting and repairing code defects, enhancing the generalization ability and detection accuracy of models is a key challenge. This paper proposes a data fine-tuning method based on Chain of Thought (CoT) fine-tuning to improve the capabilities of models on defect detection in AI code. We constructed a dataset that includes the CrossVul dataset and a manually created dataset of AI code defects and repairs, improving data quality through techniques like context free removal. In the experiments, we used the Codeshell-7B, Qwencoder2.5-7B and Llama3.1-7B as the base and trained them using LoRA fine-tuning techniques. We compared different datasets and training methods to verify the model's effectiveness in detecting and repairing AI code defects. The results show that the CoT fine-tuning model outperforms models without CoT fine-tuning in all aspects of handling code defect tasks. Additionally, the specialized dataset we created for AI code defects and repairs significantly enhances the model's accuracy and repair rate in AI code detection. Our experiments highlight the importance of constructing targeted datasets for AI code defects and employing CoT fine-tuning strategies in improving code defect detection.
The rapid rise of malware challenges traditional detection methods due to code obfuscation and polymorphism. While machine learning classifiers offer quick detection and can identify complex malicious features, they are susceptible to backdoor attacks. We introduce GAT, a genetic algorithm-based approach for generating effective and stealthy Android backdoors. Using the SHAP interpretability tool, we first select efficient features as primary backdoors. A fitness function then enables iterative optimization through a genetic algorithm. Additionally, we propose a method to integrate backdoor features into the source code, maintaining functionality while facilitating attacks on Android classifiers in real data outsourcing scenarios. Our evaluation of the Drebin and Mamadroid malware detectors in data outsourcing scenarios indicates that an attack success rate exceeding 70
