31 August 2026, Volume 4 Issue 4
-
Select all|
-
Journal of Cybersecurity. 2026, 4(4): 2-16. https://doi.org/10.20172/j.issn.2097-3136.260802
Audio large language models have been applied in various scenarios including voice assistants, intelligent customer services and human-computer interaction. Their training, fine-tuning and data construction processes often rely on third-party platforms or external data providers, which may introduce backdoor risks through the outsourcing of data and computing resources. Existing studies have demonstrated that audio large language models exhibit significant biases in understanding different dialects, which can be exploited as a novel backdoor attack surface. To address the new security threats arising from dialect differences, this paper investigates backdoor attacks on audio large language models triggered by dialect features. Taking target dialect speech as the trigger, the proposed method constructs backdoor samples via audio data poisoning and speech synthesis. The trained audio large language model can produce malicious responses once receiving target dialect inputs, thereby enabling dialect-driven targeted discrimination attacks. We conduct experimental evaluations on multiple datasets and mainstream audio large language models. The experimental results show that our method achieves high attack success rates across diverse task scenarios while imposing limited impact on model usability for clean speech inputs. The research findings verify the feasibility and severe practical hazards of dialect-triggered backdoor attacks, and reveal potential security risks of audio large language models in the scenarios of outsourced training, data poisoning and natural language variant input.
-
Journal of Cybersecurity. 2026, 4(4): 17-28. https://doi.org/10.20172/j.issn.2097-3136.260805
Existing deepfake video detection methods mostly implement only binary classification and lack supporting detection evidence. To solve this problem, this paper studies an interpretable detection method based on key visual evidence. RGB frames and optical flow maps were extracted from multiple consecutive clips of input videos. A front-end detection model was adopted to identify video authenticity, and the Grad-CAM algorithm was used to generate heatmaps corresponding to the RGB branch and optical flow branch, which characterize static forgery traces, motion inconsistency and core attention regions of the model. On this basis, representative key visual evidence was screened according to frame-level prediction scores and dual-branch fusion scores. Four types of visual inputs, including RGB frames, RGB branch heatmaps, optical flow maps and optical flow branch heatmaps, were applied to support structured explanation generation and explanation model fine-tuning. To balance explanation quality and deployment cost, this paper constructs a lightweight interpretable framework with high-capability model for annotation construction and lightweight model for explanation generation. Specifically, ChatGPT-5.4-High with powerful visual understanding ability was used to generate initial structured explanations based on key visual evidence and video ground-truth labels. After manual verification and screening, high-quality structured annotation datasets were formed. Taking the annotations as supervision signals, the Qwen-3.5-4B model was fine-tuned via the LoRA method. The optimized lightweight model can generate standardized, evidence-consistent and logically coherent structured explanations at low computational costs according to detection results and key visual evidence. Experimental results show that the proposed method maintains the decision independence of the front-end detection model and achieves an explanation consistency score of 0.876. User evaluations demonstrate excellent performance of the model in accuracy, completeness and comprehensibility, with scores ranging from 4.75 to 4.89 (full score: 5). The proposed method can provide a reliable basis for the analysis and review of deepfake video detection results.
-
Journal of Cybersecurity. 2026, 4(4): 29-41. https://doi.org/10.20172/j.issn.2097-3136.260806
Graph-based abnormal behavior detection is often challenged by noisy neighborhoods, class imbalance, and insufficient reliability characterization of prediction results. In fraud-related graphs, abnormal nodes are usually sparse, and their discriminative signals may be weakened by surrounding normal nodes during local message passing. To address these issues, this paper proposes a conformal abnormal behavior detection method based on a self-representation graph neural network. The model first learns node representations from attributes and graph topology through a graph neural encoder, and produces initial prediction results. Then, a kernel self-representation module is introduced in the embedding space to capture cross-neighborhood reconstruction relationships among nodes. The reconstructed information is fused with the original embeddings for supervised classification, thereby improving the separability of abnormal nodes. Meanwhile, reconstruction coefficients and residuals provide auxiliary structural evidence for abnormal judgment. To further improve the reliability of model decision-making, conformal calibration is conducted on an independent calibration set. By jointly considering classification confidence and reconstruction consistency, a structure-aware nonconformity score is constructed to generate prediction sets with coverage guarantees and uncertainty quantification information. This design supports high-confidence abnormal alerts, suspicious-sample review, and low-risk sample filtering, improving the reliability and operability of abnormal detection results. Experiments on public abnormal behavior graph datasets show that the proposed method improves abnormal-class recognition and provides reliable and interpretable outputs for review fraud detection, illicit transaction identification, and abnormal account analysis.
-
Journal of Cybersecurity. 2026, 4(4): 42-52. https://doi.org/10.20172/j.issn.2097-3136.260523
Generative text steganography mainly relies on language models to generate steganographic texts. However, existing methods are generally trained on a single corpus and generate steganographic texts with random lengths, which cannot adapt to dynamically changing message types and channel conditions. To solve this problem, this paper proposes a large language model-assisted intelligent text steganographic transmission method to effectively reduce the transmission overhead. In this method, a large language model is fine-tuned with multiple corpora in a parameter-efficient manner to obtain adapters applicable to different text styles. Based on the token probability distribution output by the fine-tuned model, secret information is recursively embedded via adaptive dynamic grouping. Furthermore, a dual-agent reinforcement learning framework is designed to adaptively optimize steganographic strategies including corpus type, output sentence length and transmission power, according to real-time channel states and security indicators such as attack success rate. This method comprehensively considers steganographic security requirements and actual transmission environments, and constructs a reward function to evaluate the performance of different strategies. The function can guide the selection of steganographic strategies and transmission power, so as to achieve a dynamic balance between concealment and transmission robustness. Experimental results show that compared with the transmission scheme based on adaptive dynamic grouping algorithm, the proposed method reduces transmission delay by 38.5% and the attack success rate to 0.83%, while maintaining basically consistent embedding rate and information divergence. Therefore, the proposed method can realize an optimal balance between concealment and transmission robustness under dynamic channel conditions and diverse message types.
-
Journal of Cybersecurity. 2026, 4(4): 53-63. https://doi.org/10.20172/j.issn.2097-3136.260808
360° video transmission provides users with immersive and interactive panoramic viewing experiences. However, it suffers from high bandwidth consumption, large transmission latency and wireless link malicious jamming, which seriously degrades users’ quality of experience (QoE). To solve the above problems, this paper proposed a reinforcement learning-based edge-assisted anti-jamming transmission scheme for 360° video. Relying on edge-side user viewport prediction information, the scheme jointly optimized tile transmission sequence, per-tile power allocation and bitrate selection to adapt to the dynamic changes of jamming, user viewport and network environment. A lightweight multi-output pointer network was designed to encode tile state features including tile spatial location, channel gain and jamming intensity. The network adopted an attention mechanism to quantify the importance of different tiles and realize the joint optimization of anti-jamming transmission strategies. This paper constructed a 360° video anti-jamming transmission game model, analyzed the existence conditions of Stackelberg equilibrium, and derived the upper bound of users’ QoE performance. Simulation results indicated that compared with baseline schemes, the proposed scheme increased the viewport QoE by 36.2% and reduced power consumption by 15.3%, which verifies the robustness and effectiveness of the proposed scheme under dynamic variations of jamming, viewport and network environment.
-
Journal of Cybersecurity. 2026, 4(4): 64-73. https://doi.org/10.20172/j.issn.2097-3136.260506
Automotive millimeter-wave radar has been widely applied in autonomous driving due to its low cost and strong robustness to adverse conditions such as illumination and smoke. However, with the increasing number of radars on roads and the growing scarcity of spectrum resources (77~81GHz), mutual interference among these radars has become a critical issue. The resulting ghost targets degrade radar sensitivity and significantly increase the false alarm rate, posing a serious threat to autonomous driving safety. To address this problem, this paper proposes a multi-domain joint ghost target removal method integrating angle of departure (AoD), angle of arrival (AoA), and Doppler information. A key finding is that interference signals have asymmetric transmission characteristics compared with target echoes: target echoes involve round-trip propagation, whereas interference involves one-way propagation. Initially, this paper performs joint angle estimation in the AoD-AoA domains. By exploiting the absence of AoD information in one-way interference signals, this paper achieves the separation of interference and target echoes in the angular domain. For ghost targets near the zero-degree angle, this paper introduces pseudo-random coding in the slow time domain to convert them into the Doppler domain, where they are effectively removed through an optimization model designed to maximize the signal-to-interference-plus-noise ratio (SINR). Experimental results demonstrate that the proposed method can effectively remove ghost targets caused by inter-radar interference, thereby enhancing the perceptual reliability of radars in complex electromagnetic environments.
-
Journal of Cybersecurity. 2026, 4(4): 74-83. https://doi.org/10.20172/j.issn.2097-3136.260546
Deep learning technology is advancing rapidly, and deep neural network models are widely used in numerous fields such as image recognition and natural language processing. However, the training and deployment of deep neural network models rely on substantial data and computational resources, leading to high costs, and their intellectual property protection has become increasingly prominent. A dual-verification copyright protection algorithm for deep neural network models based on adversarial examples is proposed. Specifically, the algorithm generates adversarial samples using the Momentum Iterative-based Fast Gradient Sign Method (MI-FGSM), embeds watermark information into these samples, and retrains the model accordingly. Through dual verification of adversarial and watermarked samples, an effective copyright authentication mechanism is established. This approach requires no modification to the model's parameters or structure; it can be implemented merely through data processing, featuring simplicity, convenience, and strong robustness. Experimental results demonstrate that the proposed method exhibits strong robustness against Gaussian filtering, cropping, and compression attacks.
-
Journal of Cybersecurity. 2026, 4(4): 84-98. https://doi.org/10.20172/j.issn.2097-3136.260531
Website hidden link detection faces significant challenges, as traditional methods often suffer from limited model generalization, heavy reliance on feature engineering, and poor interpretability of detection results. To better address these issues, this study proposes a zero-shot hidden link detection method driven by chain-of-thought prompting in large language models. By leveraging the strong natural language understanding and complex reasoning capabilities of large language models and chain-of-thought prompting strategies specifically designed for hidden link detection scenarios, the proposed method effectively enhances model generalization and innovatively provides interpretable outputs for detection results. Both subjective and objective metrics are employed to quantitatively evaluate the performance of the large language model and to verify the effectiveness of the prompting strategy in hidden link detection. The evaluation metrics include accuracy, precision, recall, and F1. Experimental results show that the proposed method achieves an F1 of 0.98, outperforming the baseline methods. This study provides an effective technical solution for hidden link detection in scenarios lacking high-quality datasets and further lays a foundation for the application of general-purpose large language models in cybersecurity detection tasks.
-
Journal of Cybersecurity. 2026, 4(4): 99-113. https://doi.org/10.20172/j.issn.2097-3136.260810
This paper proposes a residual neural network-based construction method for differential-linear distinguishers for the lightweight block cipher Simeck32/64, and conducts corresponding key recovery attacks. The proposed method combines the advantages of differential cryptanalysis and linear cryptanalysis, and designs a novel input data format containing both ciphertext pair difference values and linear mask components. A deep residual one-dimensional convolutional neural network model is constructed to capture the complex statistical biases generated in the multi-round encryption process of the cryptographic algorithm. Aggregating multiple ciphertext pairs into a single sample enhances the model’s ability to perceive weak statistical features. Compared with existing neural distinguishers based solely on differential analysis, the proposed method achieves higher classification accuracy with lower data complexity for 7 to 10 encryption rounds. Based on the constructed distinguisher, key recovery attacks on 9 to 13 rounds of the Simeck32/64 are implemented. The experimental results verify the effectiveness and feasibility of the differential-linear analysis method in neural network-assisted cryptanalysis.
Volume 4 Issue 4
Bimonthly
Sponsored by China Aerospace Academy of Systems Science and Engineering
ISSN 2097-3136
CN 10-1901/TP
-
Total visitors
Visitors of today
Now online