The Linux kernel is a highly configurable system with over 32,000 options, enabling deployment across a wide range of platforms, from mobile devices to supercomputers. However, this configurability introduces substantial complexity: misconfigurations frequently lead to build failures, including compilation failures and boot-time panics, both of which are notoriously difficult to diagnose and resolve. Existing approaches primarily address compilation failures and provide little support for repairing boot-time panics. Furthermore, the lack of large-scale error configuration datasets limits the ability to rigorously evaluate the effectiveness of repair techniques. Another critical challenge is avoiding negative side: a fix for one failure can inadvertently introduce new failures elsewhere in the configuration. In this paper, we present KconfigHealer, the first framework that leverages Large Language Models (LLMs) to automatically repair build failures based on the following two innovations: 1) Our approach integrates prompt engineering with system-level feedback and introduces ConFixer, a dynamic repair mechanism designed to preserve semantic consistency and suppress cascading failures. 2) We introduce ConfigDestroyer, a novel tool for systematic configuration failure injection, and release a dataset of 108,075 failures (kernel versions 4.19, 5.4, 5.10, and 6.6.) generated using this approach - an order of magnitude larger than prior datasets. Experimental results show that KconfigHealer can repair build failures in an average of 3.70 iterations, demonstrating significant advantages over SOTA methods, particularly in handling transitive dependency breaks and cross-module coupling conflicts.
Dynamic Searchable Symmetric Encryption (DSSE) reconciles data confidentiality and usability in outsourced storage systems by enabling keyword search and continuous data update over encrypted data. Conjunctive DSSE, as an expressive variant of DSSE, allows clients to retrieve data matching multiple keywords simultaneously. All existing conjunctive DSSE schemes, as we know, assume that the secret-key remains perfectly secure; however, key compromise incidents occur frequently in practice. Once the secret-key is compromised, existing conjunctive DSSE schemes lose their privacy guarantees entirely. To address this issue, we define a new framework for conjunctive DSSE with key update and introduce post-compromise security under the paradigm of leakage function. Within this framework, we develop ${\sf Poseidon}$, the first conjunctive DSSE scheme that achieves post-compromise security. We provide a formal security proof showing that ${\sf Poseidon}$ preserves both data confidentiality and operation privacy even if the secret-key is leaked to the adversary. We conduct an extensive performance assessment of ${\sf Poseidon}$ on real-world datasets and benchmark it against current conjunctive DSSE schemes under the same security level. The results show that ${\sf Poseidon}$ achieves 1.58$\sim 4.99\times$ speedup in search time beyond the prior and fastest scheme and incurs comparable communication overhead and update latency.
Image steganography, as one of the key technologies in digital communication, aims to achieve the secure exchange of secret information by imperceptibly embedding it into cover images. However, existing steganographic methods still face significant limitations in terms of security and visual quality. To address these challenges, this paper proposes a high-security image steganography model based on an improved Transformer architecture (AGDNet). By integrating innovative multi-scale feature convolution modules and a composite loss function design, the proposed model makes significant advancements in enhancing both the security and visual quality of steganographic information. Experimental results demonstrate that the proposed method exhibits superior performance across several key metrics, such as PSNR, SSIM, MAE, and RMSE, fully validating the effectiveness and advantages of AGDNet. In experiments conducted on the DIV2K dataset, the PSNR values for secret/recovered image pairs and cover/stego image pairs reached 58.80 and 57.64, respectively, outperforming existing state-of the-art steganographic methods. Moreover, AGDNet shows ex ceptional security in anti-steganalysis experiments, particularly when tested against five deep learning-based steganalysis models, where it significantly outperforms traditional methods. The experimental results indicate that AGDNet effectively enhances the concealment and security of steganographic information. https://github.com/SSSunShuaiSS/AGDNet
The efficacy of coverage-guided library fuzzing is strictly bound by the quality of harnesses that map unstructured inputs to semantically valid API sequences. While recent Large Language Model (LLM)-based approaches have reduced the manual burden of harness construction, they frequently suffer from a lack of domain context, yielding harnesses that fail to compile or violate implicit usage constraints. Our preliminary work, PromeFuzz (ACM CCS 2025), addressed this challenge by injecting statically extracted metadata into LLM prompts to ensure initial syntactic and semantic correctness. However, longitudinal analysis reveals a critical coverage plateau: statistical improvements in API coverage do not translate proportionally to branch coverage due to deep-seated Fuzzer Blockers, such as complex input predicates and state-dependent execution prerequisites. To transcend this limitation, we present ProteusFuzz, an adaptive, closed-loop framework that synergizes static knowledge-driven synthesis with dynamic, diagnosis-driven remediation. We propose a novel empirical taxonomy of blockers, distinguishing between seed-related barriers (requiring format-compliant inputs) and harness-related barriers (requiring state-aware driver logic). Consequently, ProteusFuzz employs a dual-path remediation strategy, leveraging an intelligent agent to orchestrate predicate-aware seed synthesis and automated harness refinement. Extensive evaluation on 20 widely used C/C++ libraries demonstrates that ProteusFuzz achieves average branch coverage improvements of 16.18% over PromeFuzz, 27.24% over PromptFuzz, and 16.80% over CKGFuzzer, while further outperforming OSS-Fuzz-Gen and OSS-Fuzz by 43.12% and 32.10%, respectively. Furthermore, it uncovered 21 previously unknown bugs, including 15 confirmed or fixed bugs, 2 of which have been assigned CVE IDs. These results underscore that bridging static correctness with dynamic adaptability is essential for advancing the autonomy and depth of library fuzzing.
Mobile Edge Computing (MEC) enables resource-constrained IoT devices to offload computationally intensive tasks to MEC servers. However, in practice, MEC servers are often regarded as semi-trusted, as they are deployed at the network edge, where they may operate in less controlled environments (e.g., base stations or roadside units), leading to higher exposure to physical tampering, limited redundancy, and potentially lower reliability compared to centralized cloud servers. In this paper, we propose the forward-secure and lightweight Authenticated Key Agreement (AKA) protocol for resource-constrained devices against semi-trusted servers. In our protocol, the MEC servers are functionally categorized into Authentication Servers (AS) and Computation Servers (CS). Our protocol ensures that the CS cannot access any user privacy-related information. We propose a security model to formally analyze and prove the protocol's security. Finally, compared with several representative AKA protocols, our protocol achieves a better overall balance between security and efficiency, offering stronger security guarantees while maintaining competitive computational and communication performance.
Past successes in applying machine learning to data provenance graphs (DPGs) – a structured audit log representation that describes the history of a system's execution – to detect intrusions have spurred growing interest in the security community. Although many studies favor intuitive graph neural network (GNN) approaches for analyzing DPGs, recent research increasingly suggests that there is an incompatibility issue between DPGs and off-the-shelf GNNs. In this work, we address this issue by revisiting the empirical evidence of DPG–GNN incompatibility reported in prior studies. We observe that existing solutions largely rely on manually designed specifications, raising concerns about their ability to generalize to unseen scenarios. To address this limitation, we present PROVPLUG, a unified LLM-driven plugin framework that generalizes two previously separate specification-based directions, namely graph modification and training guidance, for improving the compatibility between DPGs and GNNs and enhancing the detection performance of provenance graph host intrusion detection systems (PIDSs). By combining the general reasoning capabilities of LLMs with domain-specific provenance knowledge, PROVPLUG introduces two complementary plugins that address the adaptation challenges of DPGs to GNNs and GNNs to DPGs, respectively. We evaluate PROVPLUG using eight state-of-the-art PIDSs across eight widely-used public datasets. The experimental results demonstrate that PROVPLUG can significantly improve the detection performance of these targeted PIDSs, while achieving notable superiority over all specification-based baselines.
Multi-owner multi-user encrypted databases enable each authorized user to search over data contributed by multiple parties in a privacy-preserving manner. Public-key searchable encryption (PKSE) appears to be a suitable solution but suffers from high latency due to costly public-key operations. Symmetric searchable encryption (SSE) benefits from sublinear search performance but is limited to a single-user setting. In this paper, we propose PMoMu, which combines the advantages of both PKSE and SSE to achieve multi-owner multi-user functionality with high efficiency. We first redesign key-aggregate searchable encryption (KASE) to support efficient and secure multi-owner and multi-user data access. Built on dynamic SSE, PMoMu satisfies the standards of dynamic security, including forward and backward privacy. We further design epoch-based forward privacy, which offers a trade-off between overhead and security for practical deployment. To support epoch-based backward privacy, we devise an epoch rotation technique using symmetric revocable encryption (SRE). We formally define the security model and prove security in both the static and dynamic settings. Finally, we implement our system and evaluate its efficiency with a real-world dataset.
Visual Question Answering (VQA) aims to answer natural language questions based on images. Users increasingly rely on online Vision-Language Model (VLM) services for such tasks. However, this introduces privacy risks, as service providers may infer sensitive information from image content. To address this, we propose ImgSan, an image privacy sanitization approach for online VQA. ImgSan performs region-specific obfuscation with adaptive intensity, preserving key visual information for answering while sanitizing privacy-sensitive content. It determines the semantic relevance between image regions and questions, and accordingly estimates the obfuscation intensities. For regions that are both relevant and sensitive, ImgSan applies utility-aware privacy enhancement to strengthen protection while maintaining answer accuracy. We further construct PrivacyVQA2K, a benchmark of 1,000 images with question-answer pairs and 2,002 annotated privacy-sensitive objects. Experiments show that ImgSan achieves superior balance between privacy protection and response utility compared with existing baselines.
Federated learning enables collaborative model training without transferring raw data, but practical deployments still lack a unified mechanism that simultaneously provides deployable mutual authentication, context-bound uploads, and verifiable client-side privacy protection. To address this problem, this paper presents a new LDPAKE framework that integrates OPAQUE, oblivious pseudorandom functions (OPRFs), local differential privacy (LDP), and zero-knowledge proofs into a single session-rooted protocol. The core idea is to use OPAQUE to establish a mutually authenticated session, derive a context-bound randomness anchor through an OPRF over the tuple (sid, t, ctr, ver, pid), and then let the client mix this anchor with fresh local entropy to generate the perturbation seed. In this way, the proof system no longer needs to encode the entire randomness-generation chain, but only proves that the uploaded perturbation is derived from the bound seed and satisfies the active clipping and noisebound policy. The resulting design preserves three properties in a unified manner: lightweight mutual authentication without PKI or trusted hardware, explicit binding of each upload to its session and round context, and verifiable client-side perturbation without revealing the plaintext update or the realized noise. The OPRF anchor strengthens context-bound seed provenance under a nonce-respecting execution in which the client fixes its local nonce before learning the anchor and one context counter admits at most one upload; it does not by itself prevent adaptive nonce grinding after the anchor is observed. Experimental results show practical access and upload performance under weak links and concurrent settings, while characterizing proof cost and the security of the explicitly encoded compliance relation.
As a typical crowdsourcing service in the healthcare field, medical crowdsourcing allocates tasks to large-scale professionals via edge-cloud collaborative platforms. It provides patients with health consultations and treatment plans. Nevertheless, it is non-trivial to achieve fine-grained access control for medical tasks, fine-grained trusted identification for task requesters, and flexible task selection for medical providers. Furthermore, current access control solutions cannot prevent improper access to medical tasks throughout the full lifecycle, thus making them susceptible to historical and future data breaches. To this end, this paper proposes a privacy-preserving Attribute-Based Matchmaking Encryption scheme called MatFBMC that ensures both forward and backward secrecy. It offloads massive puncture operations to the crowdsourcing platform, thereby reducing the client key rotation overhead while ensuring forward secrecy. Meanwhile, this scheme ensures backward secrecy through the embedding of the time epoch, thereby realizing lightweight privilege revocation. Furthermore, the bi-directional matches between access/identification policies and the two parties’ attributes are employed to achieve bilateral fine-grained access control. We provide security proofs for semantic security against selective Chosen Plaintext Attack and existential unforgeability under Chosen Message Attack. Based on the Java Pairing-Based Cryptographic Library and real-world medical datasets, simulations display better computational and storage performance compared to the state-of-the-art works.
Correctly fixing kernel vulnerabilities remains challenging and has the potential to introduce new security risks. To mitigate this problem, we conducted a manual analysis of 182 incorrect kernel patches and found that incorrect patches typically resulted from changes to variable read and write operations by the patch. Based on this observation, we propose KLAUS++, a novel approach for assessing the correctness of patches. KLAUS++ leverages abstract interpretation to extract patch-induced modifications to variable read and write operations in the Linux kernel. It takes advantage of these alterations as coverage and branch-resolving mechanisms to guide a kernel fuzzer toward relevant code and contexts. Furthermore, it integrates a distance-guided strategy into a hybrid feedback scheme, enabling the fuzzer to rapidly converge toward target code regions and efficiently navigate deep execution paths. We implemented and evaluated KLAUS++ on numerous real-world Linux kernel patches. So far, KLAUS++ has identified and reported 30 incorrect patches to the Linux community, some of which could enable privilege escalation on Android and Ubuntu systems. The evaluation shows that the applications of the distance-guided strategy can improve the efficiency of triggering the newly introduced vulnerabilities by 35.4%.
Federated learning has gained prominence as a privacy-preserving distributed learning paradigm, yet its environmental impact, particularly its carbon footprint, remains inadequately quantified and understood. Existing carbon accounting methods for federated learning predominantly focus on operational emissions from electricity consumption, employing oversimplified and deterministic models that neglect embodied carbon from hardware manufacturing and inherent spatiotemporal uncertainties in carbon intensity. To address these limitations, this paper introduces FLCarbon, a probabilistic carbon accounting model that holistically estimates the carbon footprint of federated learning systems by integrating both operational and embodied carbon components while explicitly modeling their uncertainties. We develop parameterized models for embodied carbon of key hardware components, System-on-Chip (SoC), memory, and storage, incorporating factors such as manufacturing yield, energy efficiency, and geographic variability. For operational carbon, we model computing and communication processes under spatially and temporally varying grid carbon intensities and network conditions. To capture parameter distributions, we construct a comprehensive dataset from diverse sources, including ESG reports, industry publications, and real-world electricity data, and apply dual-stage distribution modeling via histograms and Kernel Density Estimation (KDE). Experimental evaluations across multiple datasets (CIFAR-10, ImageNet, FEMNIST, Speech Commands, Common Voice) and federation strategies (FedAvg, FedAdam) demonstrate that FLCarbon provides distribution-aware carbon estimates, revealing significant deviations from state-of-the-art deterministic models (e.g., up to 339% higher in federated settings). These results underscore the critical importance of embodied carbon and uncertainty-aware modeling in accurately assessing federated learning's environmental impact. Our work not only advances carbon accounting for FL but also provides a foundational framework for developing truly carbon-efficient federated learning systems.
This paper addresses the safety guardrail problem of multiagent systems under AI-generated attacks. We first propose a novel attack content generation mechanism based on a large language model. This generator efficiently explores the adversarial parameter space and produces a diverse set of attack contents in a single pass, reducing the cost of attack generation. To address this threat, we propose a new optimal security control scheme for multiagent systems that models the attack-control interaction as a coupled game and develops a composite control law, thereby achieving the optimal control under AI-generated attacks. To solve the Hamilton–Jacobi–Isaacs equation, an adaptive multistep policy gradient algorithm with an actor–critic architecture is proposed. This algorithm improves learning efficiency and reduces approximation bias, while balancing convergence speed and training stability. Finally, we demonstrate the defensive capability for multiagent systems with AI-generated attacks through simulation and physical unmanned aerial vehicle experiments.
Device-to-device (D2D) communications in cellular systems improve spectral efficiency and reduce latency but also create security challenges, as suspicious D2D links can bypass base station monitoring and evade detection. This paper proposes a theoretical framework for covert surveillance of such links. A full-duplex Warden legitimately monitors a suspicious D2D pair via eavesdropping in the presence of a base station and a cellular user (Alice), covertly injecting artificial noise to aid eavesdropping. We derive the average minimum detection error probability at the D2D receiver and its optimal threshold for both underlay (spectrum reuse) and overlay (orthogonal spectrum) modes. To evaluate performance, we present an eavesdropping non-outage probability metric, capturing the reliability of the legitimate eavesdropping link while ensuring quality-of-service (QoS) for cellular and D2D communications. To maximize this metric, the transmit powers and the spectrum partition factor are optimized under both modes, considering a fixed-power adversary and an adaptive one that adjusts power to maintain QoS. We solve the resulting non-concave problems by a unified concave-convex procedure. Additionally, a flexible mode selection scheme is proposed to reduce selection overhead while meeting a minimum surveillance requirement. Extensive numerical results demonstrate the effectiveness of the proposed framework in enhancing covert surveillance.
Imagine a group of users wishing to perform offline micropayments (e.g., splitting the daily dinner bills) without revealing who paid what or relying on a central coordinator. Existing blockchain Payment Channel Networks (PCNs) fall short in such scenarios: they are designed for pairwise payments instead of multi-party settings, leak transaction metadata, and struggle to ensure consistent channel state under adversarial conditions. We present MystChannel, a privacy-preserving blockchain multi-party payment channel protocol. MystChannel introduces Blind Ring Adaptor Signatures (BRAS) to achieve transaction unlinkability and value privacy, independent of fixed payment amounts. To further ensure robustness against malicious par ticipants, we integrate an aggregable multi-signature scheme with a pipelined multi-leader Byzantine Fault Tolerant consensus protocol, enabling concurrent transaction processing and con sistent state agreement. MystChannel is provably secure under the Universal Composability (UC) framework. We implement and evaluate MystChannel, demonstrating that it achieves a 59.77× speedup over Bitcoin on-chain payments and is 77.375× faster than the state-of-the-art privacy-preserving multi-party channel AMHL [NDSS’19]. Our results affirm MystChannel as an efficient and secure off-chain payment infrastructure.
Secure data transmissions in open maritime environments present significant challenges due to dynamic topologies and privacy requirements. Current identity-based schemes often lack context awareness and adaptation. To address these limitations, an anonymous authenticated group key agreement with context-aware identity management is proposed. The method models identity validity as a context-dependent semantic state governed by attribute–membership–region constraints. First, a context-aware identity management mechanism is established under a contextual legitimacy policy. Second, a decentralized authentication protocol combines pseudonymized ring signatures with threshold-based revocation. Finally, a dual-layer key evolution mechanism supports rapid session recovery and periodic group key updates without the need for trusted third-party dealers. Formal security analysis demonstrates that essential security requirements are satisfied, while performance evaluations confirm the superiority of the proposed scheme over existing approaches.
The preservation of data privacy in smart contracts emerges as a critical requirement. Privacy issues can arise after contract deployment due to coding errors or emerging risks. Our studies reveal that even if the contract is upgraded by applying encryption to exposed private data, significant exposure risks remain, as attackers may still infer sensitive information from previously public information. In this paper, we are the first to highlight the persistence of privacy leakage after contract upgrades. We provide a comprehensive characterization of this persistent privacy leakage. We then introduce a transaction-sequence-based method to mitigate persistent privacy leakage, accompanied by a source-variable-driven optimization strategy that ensures both efficiency and effectiveness. We develop PandoraGuard, the first tool designed to enhance post-deployment data privacy across a wide range of Ethereum applications. Our evaluation demonstrates the effectiveness and efficiency of PandoraGuard.
