Firmware rehosting executes firmware images in emulated environments such as QEMU to enable scalable dynamic analysis of Internet of Things (IoT) devices. In practice, rehosting pipelines remain fragile across diverse real-world firmware images, as reaching an externally observable execution state depends on tightly coupled artifacts spanning boot scripts, persistent configuration (e.g., NVRAM-like key-value state), and network setup. Template-driven frameworks often fail to accommodate long-tail vendor conventions, while unconstrained use of large language models (LLMs) risks unsupported modifications and irreproducible executions. We introduce FirmPilot, an evidence-guided multi-agent framework for environment recovery in firmware rehosting. FirmPilot reformulates rehosting as iterative environment reconstruction in which a search agent grounds decisions through similarity-based retrieval, a planner coordinates execution-accepted transitions, and specialized agents recover filesystem/init artifacts, persistent state, and network exposure. Through repeated execution and evidence-grounded artifact deltas, the system resolves cross-layer dependencies across boot, state, and networking that otherwise prevent firmware executions from reaching a stable, externally reachable state in emulation. Evaluated on the large-scale, real-world LFwC firmware corpus, FirmPilot improves web-service reachability over FirmAE from 25.49

The number theoretic transform (NTT) is a core building block of lattice-based post-quantum cryptography and a major target for soft analytical side-channel attacks (SASCA), especially in Internet of Things (IoT) deployments where ML-KEM runs on resource-constrained microcontrollers with stable physical leakage and tight performance and memory budgets. Existing SASCA attacks against ML-KEM usually incur high computational cost and depend heavily on leakage from the inputs to the first NTT layer. To address these limitations, we propose GRALE, a practical SASCA scheme for single-trace attacks on the NTT during ML-KEM key generation. GRALE combines graph-based reduction (GR) and auxiliary leakage exploitation (ALE). GR removes the first few NTT layers from the factor graph by exploiting the invertibility of their values under ML-KEM’s small-secret distribution, reducing graph size and message-passing cost. ALE further exploits auxiliary leakage from more informative side-channel observations in subsequent polynomial multiplications in the NTT domain, reducing reliance on first-layer input leakage. Under a Hamming-weight leakage model with additive Gaussian noise, and using measurements from a Cortex-M4 microcontroller, GRALE achieves full key recovery when noise satisfies σ < 0.6, reduces end-to-end attack time by up to 50% compared with state-of-the-art baselines, and gains a further 30% improvement when ALE is combined with GR.

Large Language Models (LLMs) have shown promising performance in software vulnerability detection, particularly after domain-specific Supervised Fine-Tuning (SFT). However, it remains unclear whether these models genuinely internalize vulnerability root causes or merely exploit surface-level functional patterns. While prior work documented related failures on pre-trained or zero-shot models, the SFT process itself, and how explicit reasoning supervision modulates it, remains under-explored. We study fine-tuned decoder-only LLMs under vanilla SFT and SFT with reasoning supervision, identifying a failure mode we term the Semantic Trap, characterized by three symptoms: pairing-sensitive performance, gap-dictated decisions, and fragility to semantic-preserving changes. To probe this, we propose TrapEval, an evaluation framework comprising two real-world datasets, V2P (vulnerable paired with patched code) and V2N (vulnerable paired with unrelated normal code), alongside semantic perturbations, CodeBLEU-based gap analysis, and an LLM-assisted reasoning failure taxonomy. Evaluating five representative LLMs fine-tuned with and without explicit reasoning (Chain-of-Thought), our results show vanilla SFT yields deceptively high scores on unpaired data (V2N) while failing all three symptoms. Models suffer high false-positive rates on V2P, degrade under perturbations, and exhibit a systematic dependency on the textual gap between vulnerable and patched code. Finetuning with explicit reasoning reduces these symptoms but costs recall; its lack of measurable gap-dependency partly reflects a floor effect rather than escaping the trap. Furthermore, our taxonomy reveals these models still misinterpret control flow and hallucinate API behavior, indicating current fine-tuning mitigates but does not eliminate reliance on surface features.

Side-channel techniques, such as Distance-of-Means and Kolmogorov-Smirnov analysis, provide valuable insights about leakages from the indirect perspectives of statistical moment and cumulative distribution function (CDF), circumventing the direct and costly estimation of leakage probability densities and thus enabling broad application. Though both the perspectives are informative, their relationship remains unclear – the question of ”which one is better and under what circumstances?” is left open. In this paper, we introduce the probability-probability (PP) plot as a common framework for explaining the mathematical foundations of the CDF-based techniques, facilitating an intuitive understanding of different variant strategies. Then, we propose a novel distinguisher based on the Mann-Kendall test, where the key identification task is reformulated as a goodness-of-fit test checking whether a key-dependent sequence originates from a random binomial distribution. Compared to the existing methods, it delivers substantial gains within the CDF-based family. Finally, we explore the symmetry and dual counterpart of CDF in mathematics, developing an interesting technique based on the inverse cumulative distribution function (ICDF). We present a general discussion of its bridging role and, on this basis, establish the relationships among moment-based, ICDF-based, and CDF-based techniques, thereby enabling the evaluation of CDF-based techniques using metrics originally proposed for the moment-based family. Our theoretical analysis accounts for the empirical observations in practice, and further shows that the improvement achieved by the Mann–Kendall distinguisher is near-optimal.

Side-channel evaluations of cryptographic implementations have been mandated at high security levels. Leakage assessments and attack evaluations are two fundamental aspects. However, there exists a mismatch between leakage assessment outcomes and attack evaluation results. Specifically, popular leakage assessment methods employ well-known statistical tools such as Mutual Information (MI) to quantify the amount of information leaked. Yet, no evidence is provided to further address the more critical problem of whether the theoretically assessed leakages are exploitable in a concrete attack. As is now widely recognized, statistical guidelines offer limited interpretability for resistance against specific attacks. This work bridges this long-standing gap through three major contributions. First, we leverage a family of globally convex/concave and coercive functions to profile a novel noisy leakage model that is grounded in geometric theory. This model enables leakage assessments within an interpretable geometric paradigm. Second, we derive multi-order leakage evidence from the model and demonstrate its exploitability in developing attacks. This is also an evaluation of leakage exploitability. We propose three novel evidence-based distinguishers to validate our claimed exploitability. Third, we formulate worst-case leakage assessments as constrained problems and develop a convergent algorithm to approximate the first-order worst-case leakage bound as an example. The improved distinguisher is derived based on the worst-case evidence. Theoretical analysis has demonstrated that the multi-order leakage evidence is consistently achievable and uniquely characterizable. Furthermore, the convergence rate toward approximating the first-order worst-case bound is ensured to be at least sublinear. Experimental analyses validate the theory, confirming its effectiveness in practical applicability.

Understanding video streaming traffic behavior under diverse Quality of Service (QoS) conditions is essential for improving both network performance and user experience. We present VQFlow, a large-scale and publicly available dataset that systematically captures encrypted video streaming traffic under controlled variations of key QoS parameters, including resolution, frame rate, and bitrate control policies. Unlike prior datasets, VQFlow combines packet-level traces with temporally aligned ground-truth annotations of playback configurations and network conditions, enabling fine-grained analysis of how QoS settings influence traffic dynamics. We demonstrate the utility of VQFlow through comprehensive characterization studies and standardized benchmarks on tasks such as QoS classification, bitrate estimation, traffic pattern modeling, and anomaly detection. Results show that VQFlow offers a reproducible foundation for research in multimedia systems, network management, and machine learning on encrypted traffic, and we release the dataset along with baseline implementations to accelerate future work in this domain.

Accurate network traffic classification is critical for defending against evolving cyber threats. However, mainstream methods relying solely on intra-flow features fail to distinguish malicious traffic with highly similar legitimate patterns, lacking contextual modeling and robustness to adversarial evasion. To address these limitations, this paper proposes GraphWave, a dynamic context-adaptive multimodal framework built on a novel heterogeneous Graph2Seq paradigm. It constructs maximum connected subgraphs leveraging attacker, target and temporal context to capture key contextual correlations of attack chains. It integrates wavelet-enhanced dynamic graph attention networks for spatial context learning and Transformer encoders for long-range intra-flow temporal modeling. A multimodal cross-attention fusion mechanism aligns spatial and temporal representations to enhance discriminative feature integration. Extensive evaluations on six real-world datasets show GraphWave achieves a 99.13% F1-score, outperforming state-of-the-art methods by 7.62% on average. Theoretical analysis and empirical results validate its strong robustness against traffic obfuscation, temporal confusion, high intra-flow similarity, and low-and-slow evasion, demonstrating its superiority in complex adversarial network environments.

Enterprise networks face an ever-growing threat from various unknown cyberattacks originating from the Internet. Anomaly-based Network Intrusion Detection Systems (NIDS) have become essential for safeguarding corporate networks, but existing technologies have some limitations in practice. (i) The interweaving of traffic from different service types increases the difficulty of identification. (ii) The scarcity of labeled malicious samples hinders the detection of both known and emerging threats. To tackle these issues, we propose a service-oriented approach to model benign traffic and leverage meta-learning to construct a robust metric space for precise sample comparison. Based on this, we develop SOFA, a two-stage traffic detection framework. In the first stage, SOFA identifies distinct service types within the network and trains independent one-class models for each, ensuring service-specific attack traffic is detected with high precision. The second stage employs a ResNet-based Siamese network to overcome the limitations posed by scarce malicious samples, enhancing detection of both known and emerging attacks. We evaluate SOFA on four widely used network tracing datasets, demonstrating that it achieves state-of-the-art performance and significantly outperforms a diverse set of existing methods in fine-grained attack detection. These results highlight the potential of SOFA for improving network security in a highly dynamic and unbalanced threat environment.

By introducing collision information, the existing side-channel correlation-enhanced collision attacks (CECAs) performed collision-chain detection, quickly filtered out candidates unsatisfying collision conditions, and extracted a part of optimal candidates for further processing, thereby rapidly and significantly reducing the key candidate space and the difficulty of key recovery. However, they are still limited by disadvantages such as serial implementation, complex parameter settings, and lack of intelligence, resulting in a low success rate (SR) of key recovery. To address these issues, we first present a collision detection framework with genetic algorithm (CDGA), which exploits genetic algorithm to detect the collision chains and has a strong capability of global searching. Second, we theoretically analyze the performance of CECA and bound the searching depth of its output candidate vectors with a confidence level using a data-driven hypothesis test that provides confidence bounds for Gaussian leakages and an approximation based on central limit theory (CLT) for non-Gaussian cases, which facilitates effective and stable population initialization. Third, benefiting from our hypothesis-test-guided design, we propose a goal-directed mutation that prioritizes promising collision candidates, thus improving efficiency and adaptability of the CDGA. Finally, to optimize the evolution of CDGA, we introduce a roulette selection strategy to employ a probability assignment based on individual fitness values to guarantee the preferential selection of superior genes. Comprehensive experiments on DPA Contest v4.1 [AES-256 with Rotated S-boxes Masking (RSM)] and an AT89S52 AES-128 platform demonstrate that CDGA achieves faster convergence and higher key-recovery SRs compared with TOC/FTC/FCC and Wiemers' cumulative-correlation selection.

Time-Triggered Communication is a key technology for Industrial Internet of Things (IIoT), 5G/6G ultra-reliable and low-latency communication (URLLC) and Vehicle-to-Everything (V2X).We name the data over the time-triggered communications time-triggered flows. Each flow consists of periodic packets that need to be delivered from a source to a destination node. Each packet needs to reach its destination before its deadline. Different flows can have different cycle lengths. To achieve assured transmission of time-triggered flows, existing efforts constrain the packets of a flow to be cyclically transmitted along the same path based on the flow’s given cycle length. Under such Fixed Cyclic Scheduling (FCS), reservation for flows with different cycle lengths can become incompatible over a shared link, limiting the total number of admissible flows. Considering the cycle lengths of different flows, a hypercycle has length equal to their least common multiple (LCM). It determines the time duration over which the scheduling compatibility of the different flows can be checked. In this work, we propose a more flexible scheduling scheme called the Hypercycle-level Flexible Scheduling (HFS) scheme, where a flow’s resource reservation can change over its different cycles within the same hypercycle. HFS can significantly increase the number of admitted flows by providing more scheduling options while remaining perfectly compatible with existing time-triggered system. We show that, (theoretically) the possible capacity gain provided by HFS over FCS can be unbounded. We formulate the time-triggered joint pathfinding and scheduling problem under HFS as an integer linear programming problem which we prove to be NP-Hard. To solve HFS efficiently, we further propose a least-load-first heuristic (HFS-LLF), solving HFS as a sequence of shortest path problems. Extensive study under real-world settings shows that HFS admits 1.5× the number of flows scheduled by FCS. Moreover, our proposed HFS-LLF can run 104× faster than solving HFS using a generic solver.

Mixture-of-Experts (MoE) architectures enable scalable and efficient large language models (LLMs) by selectively activating expert sub-networks through a routing mechanism. However, this adaptive design introduces a new attack surface: specific experts become disproportionately correlated with certain tokens (e.g., end-of-sequence), allowing adversaries to manipulate model behavior via lightweight perturbations. In this work, we present Groundhog Bit-Flip Attack (GBFA), the first bit-flip-based Denial-of-Wallet availability attack against MoE-based LLMs. By identifying and flipping routing-layer bits associated with related expert activations, we demonstrate that GBFA substantially extends the decoding token usage across three different LLM modes: conversational, reasoning, and agentic tasks, while largely preserving semantic fidelity. Across four main real-world MoE-based LLMs, manually deactivating on average fewer than 4 experts drives average output inflation to 5912%, with the majority of test samples reaching max tokens. These results reveal a robustness vulnerability of MoE architectures to bit flip, and highlight the potential of GBFA as an availability attack against LLMs.

Recent studies show that deep neural networks are extremely vulnerable, especially vulnerable to adversarial examples in image classification models. However, existing defenses suffer from limited adaptability across attacks, an unfavorable trade-off between clean accuracy and robustness, and substantial training-time overhead. To tackle these problems, we present a novel component, named the 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 two popular datasets. The empirical evaluations show that our scheme realizes significant outcomes against various attacks with negligible additional training overhead, while hardly degrading clean-sample accuracy.

In real-world classification tasks, the classes of data are not immutable, especially as new classes appear over time. Enabling the model to detect emerging classes in streaming data is an important research hotspot. It requires that the model can recognize new/known classes and be updated in an incremental manner. In this paper, we propose a novel learning paradigm Portray Learning to cope with this problem. So-called Portray Learning refers to maintaining a series of portrayers (e.g., AutoEncoder) where each portrayer fits a category of data. Our scheme is different from typical class-incremental learning and semi-supervised learning. The core idea of Portray Learning is transforming known/new class identification issues into multiple independent anomaly detection problems. Each portrayer is responsible for detecting whether the instance belongs to its own class and these portrayers are independent of each other, so they can run in parallel and incrementally update portrayers of new classes. Moreover, we adopt the Extreme Value Theory to automatically determine outlier thresholds for each portrayer. The empirical evaluations involving several datasets and real-world streams show that our proposal achieves better outcomes than existing methods, even with little prior knowledge.

Masking is one of the most popular to guarantee the security of the cryptographic implementation against side-channel analysis (SCA). Low-entropy masking scheme (LEMS) has been proposed to relieve the high overhead by reducing the entropy of the mask. The masks are selected carefully to resist univariate first-order attacks. However, in practice, securely implementing masking can be a challenging task. In this work, we exhibit the vulnerability of LEMS implementation on field programmable gate array (FPGA). First, we provide the security model to characterize the masking balance at the implementation level. Based on the security model, we exhibit a first-order leakage in the netlist of the implementation. It can be proved that this defect is due to the uncontrollable electronic design automation (EDA) tool to optimize the advanced encryption standard (AES) Sbox implementation on the specific architecture of FPGA during synthesis. The discovered flaw can be exhibited by performing a couple of first-order attacks to recover the secret key successfully on FPGA. The vulnerability is verified by simulation and practical measurements. Finally, the potential countermeasure and security evaluation process for LEMS are provided to avoid leakage during the pre-silicon design phase.

Understanding cyber threats is crucial for effective defense in the field of cybersecurity. If we can automatically map Common Vulnerabilities and Exposures (CVEs) to attack tactics and techniques, it will help practitioners quickly analyze reports and take responsive actions. In this paper, we introduce CyberLLM, leveraging the tailor-made large language model for mapping CVEs to cyber threat tactics and techniques. Specifically, we model the mapping of CVE to tactics and techniques as a multi-label classification problem, given that many CVEs correspond to multiple techniques of ATT CK. Then, the text description is vectorized through the tokenization process, and we deploy a series of data augmentation techniques to enrich the semantic information. Considering that external knowledge bases are helpful to enhance the contextual information of the queried CVE, CyberLLM designs a retrieval strategy based on the Jaccard distance calculation. Finally, we support flexible model fine-tuning to adapt to the needs. Through extensive experiments, we demonstrate the superiority of CyberLLM compared with 7 representative state-of-the-art methods. We also perform ablation experiments on data augmentation and evaluate the effectiveness of using retrieval information. Furthermore, we provide a series of deep insights in terms of feature attribution and attention weight visualization.

High-throughput enterprise networks face increasingly sophisticated cyber threats, making dependable and real-time malicious traffic detection a cornerstone of cyber security and operational resilience. Existing methods suffer from two critical flaws: lightweight intra-flow approaches lack service-specific discrimination, leading to high false negatives against mimicry attacks; meanwhile, static graph-based inter-flow methods incur prohibitive latency and memory costs, failing to adapt to dynamic network environments. To address these issues, we propose RearView, a two-stage dynamic relationship-aware framework for real-time malicious traffic identification. First, it deploys service-personalized Variational Autoencoders (VAEs) to rapidly screen and filter out benign traffic. Second, for the remaining suspicious flows, RearView constructs dynamic session graphs and applies anomaly-weighted graph neural networks to uncover hidden inter-flow attack patterns. A streaming cache mechanism ensures this relational mining remains highly efficient and adaptable. Extensive evaluations on four benchmark datasets and one self-constructed CrossAttack dataset show that RearView achieves an average Weighted F1-score of 97.79%, outperforming state-of-the-art methods by 10.19% on average. It reduces the false positive rate by 4.99% and false negative rate by 34.44%, while attaining a significantly lower detection latency than contemporary methods of the same type, and exhibits robust resistance against mimicry and data pollution attacks.

Private information retrieval (PIR) allows a client to privately request a block of data from a database such that no information about the queried block is revealed to the database owner. With the rapid rise of cloud computing, data is often shared across multiple servers, making multi-server PIR a promising privacy-enhancing technology. As the demand for faster keyword PIR protocols increases, current single-server PIR schemes suffer from significant computational and communication costs, while two-server PIR schemes demonstrate superior performance in this regard. In this paper, we address the problem of the keyword PIR against some adversary who can corrupt at most one party in our protocols in the semi-honest setting. A feasible two-server scheme DPF-PIR is presented, inspired by the original employment of the distributed point function. Without the need of downloading some “hint” about the database, DPF-PIR can achieve similar throughput results with the state-of-the-art single-server scheme, SimplePIR, and 25.5× faster than previous schemes. Meanwhile, the communication cost of DPF-PIR, which exhibits logarithmic complexity, is significantly lower compared to other schemes; for example, it is less than 2

Random Delay Insertion (RDI) is one of the most investigated types of hiding countermeasure. Multifarious approaches to compromise RDI have emerged over the past two decades, with a horizontal perspective considering waveform segment as a unit. Interestingly in this paper, we transform such traditional perspective to a novel vertical one and find that the distribution feature of samples extracted vertically exposures the existence of delays directly, which can be exploited for both RDI detection and measurements alignment. On this basis, we conceive a generic approach called VCAlign, leveraging binary classification to align the measurements sample by sample vertically. This approach favors adversaries significantly since it requires no reference trace, no prior knowledge, no profiling stage. As a case study, we further propose a practical paradigm and evaluate it on several RDI-protected implementations on an ARM Cortex-M4 micro-controller. The experimental results demonstrate that VCAlign has a powerful capability to eliminate the delays completely while remaining the encryption-related segments, which can be regarded as an enhanced alignment solution to conquer RDI. We firmly believe that VCAlign is a valid application of instruction effects on physical leakage from the novel vertical perspective that could bring new vitality to the alignment solutions.