Embedded PowerShell commands or scripts are among the most popular malware payloads. For malware that prioritizes stealthiness, such as fileless malware, PowerShell’s access to Windows API functions without additional libraries makes it useful for evading detection. Detecting malicious PowerShell scripts and commands is an open challenge for proactive endpoint protection due to three major issues: (1) The malicious commands are usually hidden in a long script beyond the processing limit of typical machine learning models. (2) They are usually mixed with bulky benign scripts. (3) Script obfuscation can easily conceal their potential matching signatures. In this article, we introduce a novel model addressing these challenges. It incorporates similarity learning, sentence transformer, sliding window method, and stochastic gradient descent (SGD) classifier. Our key insight is that malicious PowerShell code, particularly when obfuscated, exhibits semantic and statistical deviations from benign administrative usage, and these deviations can be captured by contrastive sentence embeddings without the need for de-obfuscation or handcrafted features. We operate this insight through a Siamese similarity learning framework that improves robustness against Out-of-Vocabulary tokens due to unseen code obfuscation methods. The sliding window method enables the model to handle long scripts, and the SGD classifier evaluates segment-level maliciousness. Our model achieves accuracies of 99.01%, 97.59%, 98.70%, and 99.73% across multiple obfuscated and mixed script benchmarks, outperforming existing baselines by over 30% in all cases. This work demonstrates a scalable and effective strategy for robust PowerShell malware detection in real-world scenarios.
As satellite systems become a greater part of critical infrastructure, they have become a significantly more appealing target for attacks. The availability of cheap off-the-shelf radio hardware has made signal spoofing and physical layer attacks more accessible than ever to a wide range of adversaries, from hobbyists to nation-state actors. Legacy systems are particularly vulnerable due to their lack of cryptographic security, and cannot be patched to support novel security measures. In this article, we use radio transmitter fingerprinting to authenticate satellite downlinks, using characteristics of the transmitter hardware expressed as impairments on the physical layer radio signal. Our SATIQ system employs a Siamese neural network and an autoencoder to extract an efficient encoding of message headers that preserves identifying information. We focus on high sample rate fingerprinting, making device fingerprints difficult to forge without similarly high sample rate transmitting hardware. We collected 10 290 000 messages from the Iridium satellite constellation at 25 MS/s, and demonstrate that the SATIQ model trained on this data maintains performance over time without retraining, and can be used on new transmitters with no impact on performance. We analyze the system's robustness against weather and signal factors, and demonstrate its effectiveness under attack, achieving an Equal Error Rate of 0.072 and ROC AUC of 0.960. We conclude that our techniques are useful for building fingerprinting systems that are effective at authenticating satellite communication, maintain performance over time and across satellite replacement, and provide robustness against spoofing and replay by raising the required budget for attacks.
Deceptive files, often called honeyfiles, have become an established tool in cyber security. Advances in machine learning (ML) models for content generation now allow the synthesis of deceptive material automatically and at scale. Metrics to quantify honeyfile attributes are thus essential to creating and evaluating effective deceptions. The two critical aspects of honeyfiles for which metrics are useful are enticement and realism. Enticement is the ability to attract the attention of intruders or users with malicious intent. Realism measures the similarity of deceptive artefacts to the objects they mimic. In the honeyfile literature, metrics for these attributes have been proposed: the Common Token Count (CTC) [1], and Topic Semantic Matching (TSM) [2] scores for enticement, and coherence and cohesion [3] for realism. In this study, we compare these metrics to the perceptions of human users exposed to text samples in a simulated data breach scenario on a crowd-sourcing platform. We recruited participants to judge the realism and enticement of honeyfile text generated using several techniques. The main findings are: (i) for the enticement metrics, TSM is aligned with the perceived enticement (p-value<0.001), while for the CTC score, we find inconsistent and inconclusive results, and (ii) for the realism metrics, cohesion and coherence, do not consistently align with perceived realism.
In cloud-Internet of Things (IoT) healthcare systems, private medical data leakage is a serious concern as the cloud server is not fully trusted. Dynamic searchable symmetric encryption (DSSE), with necessary forward and backward privacy security properties, enables doctors to retrieve ciphertexts while guaranteeing data privacy. However, existing forward and backward private DSSE schemes are not well-suited for cloud-IoT healthcare systems with attribute-value type databases. To this end, we propose an efficient privacy-preserving conjunctive searchable encryption scheme for cloud-IoT healthcare systems, called PC-SE. It is the first conjunctive DSSE scheme designed for attribute-value type databases. Specifically, we design flexible search capabilities for PC-SE to address users' various search requirements. It can not only achieve precise conjunctive search based on keywords but also realize broad attribute search. Moreover, our scheme achieves fine-grained search for attribute values while maintaining forward and Type-I-backward privacy. This approach reduces the communication burden and minimizes the risk of privacy exposure. To ensure that users with different authorities can only access the corresponding attribute values, we introduce an attribute access control mechanism in PC-SE. Finally, security analysis and experimental results demonstrate that PC-SE is secure and effective.
The swift growth of e-commerce has led to an increase in fraudulent activities, which results in significant financial losses for both suppliers and consumers. Current research on detecting fraudulent activities within e-commerce platforms primarily focuses on analyzing individual user behavioral patterns over time or examining the spatial relationships among users. However, considering temporal or spatial contexts alone is not sufficient for fraud detection since they may not exist in real scenarios. Additionally, the issue caused by the imbalance of the data to be classified has not been solved in the field of fraud identification. To address these challenges, a novel scheme is proposed for fraudulent user detection in this work. The main contribution lies in the spatiotemporal fusion of user behavior and the layer-by-layer selective aggregation of graph models. Specifically, we utilize a long short-term memory model and a multi-layer perceptron model to extract the discriminant features from time-dependent and time-independent user behavior, respectively. This approach enhances the model’s ability to detect fraudulent users with different behavioral characteristics, including time-correlated and/or time-independent fraud behavior. Furthermore, a shared classifier is added to general graph neural network, it reclassifies the output of each layer of the graph model and reconstructs the spatial neighbor relationship. This little trick makes minority class samples select similar samples with a greater probability to build their spatial neighbor relationships, which can alleviate the issue of data imbalance. In the numerical experiments, three real datasets are used to validate the proposed scheme. Experiment results, including performance evaluation, comparison with existing benchmark approaches and ablation analysis, are presented and discussed.
The complexities introduced by compiler optimization have long stood as a significant obstacle in binary analysis and reverse engineering. Function inlining, in particular, complicates function recognition by replacing function calls with the entire body of the callee, mixing code from multiple functions. State-of-the-art approaches can identify inlined functions at basic block granularity, but cannot determine which instructions belong to each function and precisely deduce inlined boundaries. Without this information, further analyses such as decompilation cannot be performed effectively. This article presents Highliner, a novel approach that improves state-of-the-art approaches by identifying inline instances at instruction-level granularity. Highliner operates downstream of block-level detectors: given basic blocks reported by state-of-the-art approaches as belonging to a specific inlined function, it labels each instruction as Inlined or Not inlined and recovers the inlined-function boundaries. We treat the problem as a sequence tagging task typical of NLP and implement a learning-based technique involving instruction embedding and recurrent neural networks. We compile a dataset of open-source projects with different optimizations and use the DWARF debug information standard to construct labeled sequences of inline instructions. We use this dataset to train, validate, and test a sequence labeling architecture in which instructions are encoded via the pre-trained assembly language transformer PalmTree and then processed by an RNN-based classifier to produce binary predictions. When evaluated as a binary classifier, Highliner achieves an F1-score of 0.94 overall. In addition, when specifically tested on recognizing function boundaries, Highliner achieves an Accuracy of 0.82 on initial boundaries and 0.83 on final boundaries.
Deep neural networks (DNNs) for image classification remain vulnerable to adversarial perturbations–subtle input manipulations that induce catastrophic misclassifications. To address this issue, we propose the Adversarial Image Rectifier (AIR), a linguistically inspired detection and mitigation framework that enhances DNN robustness by intercepting and inverting adversarial perturbations at the feature level. Unlike existing defenses, AIR operates without prior knowledge of attack patterns: it first encodes hierarchical hidden-layer feature maps of a DNN into semantically structured sentence representations, then identifies adversarial inputs through “sentiment” anomalies in these sentences–a linguistic metaphor for subtle adversarial traces. Crucially, we pinpoint a pivotal intermediate layer where adversarial perturbations dominantly propagate and train a lightweight rectifier network to selectively nullify adversarial features at this layer while preserving benign semantics. Extensive experiments on Tiny-ImageNet, CIFAR-10, SVHN, and MS COCO demonstrate that AIR achieves a correction rate of up to 95.02% and 94.62% when defending against known attacks and unknown attacks, respectively, significantly surpassing existing defense techniques.
Recent poisoning attacks on federated learning (FL) generate malicious model updates that circumvent widely adopted Euclidean distance-based detection methods. This article proposes a new defense mechanism, namely, GradCAM-AE, against model poisoning attacks on FL, which integrates Gradient-weighted Class Activation Mapping (GradCAM) and autoencoder (AE) to offer a substantially more powerful detection capability compared to existing Euclidean distance-based approaches. Particularly, GradCAM-AE generates a heat map for each uploaded local model update, transforming each local model update into a lower-dimensional, visual representation. An AE further reprojects the GradCAM heat maps of all local module updates with improved distinguishability, thereby accentuating the hidden features of the heat maps and increasing the success rate of identifying anomalous heat maps and malicious local models. A comprehensive evaluation of the proposed GradCAM-AE framework is conducted using the CIFAR-10 and GTSRB datasets under both Independent and Identically Distributed (IID) and Non-IID settings. The ResNet-18 and MobileNetV3-Large models are tested. The results substantiate that GradCAM-AE offers superior detection rates and test accuracy of FL global model, juxtaposed with contemporary state-of-the-art methods. Our code is available at: https://github.com/jjzgeeks/GradCAM-AE.
In aviation, safety is paramount, with air traffic control (ATC) playing a crucial role in monitoring aircraft to prevent collisions and manage traffic flows. In response to increasing air traffic, a renewal process has been initiated. This includes deploying the automatic dependent surveillance-broadcast (ADS-B) communications protocol, which aims to enhance surveillance precision and increase the number of aircraft that can be handled simultaneously. This transition is transforming ATC from a radar-based system to a more advanced satellite-based global positioning system (GPS) location tracking system. However, due to its inherently open design, the ADS-B protocol lacks critical security features such as authentication, necessitating the adoption of additional security measures to mitigate potential cyber-attacks. To address these vulnerabilities, this work introduces Swarm, an innovative distributed ledger-based framework, built on top of the ADS-B protocol and aimed at enhancing the security of air traffic control (ATC) while avoiding single points of failure. Swarm can be integrated into existing ATC infrastructure without requiring any modifications to the ADS-B protocol. We evaluate Swarm through rigorous and realistic attack scenarios, using real-world aviation data, demonstrating its capability to enhance the security of the aviation domain.
In an increasingly digital and interconnected world, the need for robust network intrusion detection systems is crucial to ensure cybersecurity. This article presents a novel approach to network intrusion detection that integrates both traditional machine learning methods and advanced reinforcement learning techniques to enhance detection capabilities and accuracy. The proposed system uses Proximal Policy Optimization, a reinforcement learning algorithm, to dynamically adjust ensemble weights, thereby optimizing the contributions of base learners, such as Random Forest and CatBoost. Additionally, a Multi-layer Perceptron-based meta-learner is employed to refine the predictions, leading to an overall improvement in detection performance. The model was evaluated on five diverse datasets, including NSL-KDD, CICIDS, TON IoT, DDoS, and UNSW-NB15, achieving an average accuracy of 97.16%, and an average precision, recall, and F1-score of 97% across all datasets. The proposed work is compared with the existing state-of-the-art detection methods demonstrating its better performance in detecting both known and novel attack types. Furthermore, the integration of reinforcement learning allowed for dynamic and context-sensitive decision-making, enabling the system to handle complex attack patterns that traditional models struggle with. The training and validation results across all datasets showed rapid convergence and minimal overfitting, further supporting the model’s robustness.
Indiscriminate data poisoning attacks are highly effective against unsupervised learning. However, recent studies show that contrastive learning is also susceptible to data poisoning attacks. As a form of data poisoning attack, the attacker adds poison to the clean pre-training dataset. This article proposes IDPA, an indiscriminate data poisoning attack targeting the encoder in contrastive learning. where the attacker’s goal is to directly poison the pre-trained encoder. The feature vectors of any clean sample and the attacked sample from the attacker will exhibit high similarity, causing the downstream classifier to misclassify the clean sample as the samples designated by the attacker. Therefore, this article formulates IDPA as a dual optimization problem and defines two loss functions: the attack effectiveness loss and the model utility loss. These losses are associated with effectively poisoning the pre-trained encoder and maintaining the accuracy of the downstream classifier, respectively. During training, the attack affects the contrastive learning algorithm and predictions are made on multiple datasets. Experimental results show that the attack success rate of 92%. This article evaluates the effectiveness of IDPA on the CLIP dataset released by OpenAI, with attack success rate of 88%.
Detecting malicious network traffic in large-scale, dynamic environments presents a significant challenge due to the complexity of network relationships and the evolving nature of cyber threats. Existing graph-based and sequence-based models often fail to capture both spatial dependencies and temporal patterns effectively, resulting in suboptimal detection. This study introduces the Multi-view Graph Adaptive Network (MGAN), a novel framework that integrates multi-hop graph neural network (GNN) aggregation with transformer-based sequence modeling to address these challenges. MGAN captures long-range spatial dependencies and temporal dynamics in network traffic, enabling the detection of complex attack patterns. It incorporates Dirichlet sampling for robust neighbor selection in sparse and noisy data environments and mutual information maximization to align multi-view representations for consistency. Additionally, a multi-view attention mechanism aggregates information across different hops, balancing local and global network context. Extensive experiments on four real-world datasets demonstrate MGAN’s superiority over 7 baseline models, achieving an average F1-Score above 97%, surpassing the best baseline by 2.35%. MGAN maintains detection accuracy above 97% and remains robust under data sparsity, achieving F1-Scores over 95% even when 40% of connectivity information is removed. Under noisy conditions, MGAN retains accuracy above 93%, outperforming baselines by over 4.5%. In zero-day attack scenarios, it achieves detection rates exceeding 96% for previously unseen attack categories. MGAN also exhibits exceptional computational efficiency, processing 2,034 samples per second with a detection time of 3.00 milliseconds per sample, outperforming all competing models in both accuracy and speed.
Detecting and mitigating Denial-of-Service (DoS) attacks is crucial for ensuring the availability and security of online services. While various machine learning (ML) models have been utilized for DoS attack detection, there is a need for innovative approaches to improving their performance, especially for the more challenging multi-class detection problem. In this article, we propose adopting a cutting-edge approach called Combinatorial Fusion Analysis (CFA), which leverages a recently developed framework to combine multiple ML models for improved DoS attack detection. Our methodology involves advanced score combination, rank combination, weighted combination techniques, and the diversity strength of scoring systems. Through rigorous performance evaluations, we showcase the efficacy of the combinatorial fusion approach. Our evaluations encompass key metrics such as detection precision, recall, and F1-score, providing comprehensive insights into the interpretability and effectiveness of our approach. We highlight the challenge faced by individual models in classifying low-profiled attacks, while excelling in other attack types. To overcome this limitation, model fusion techniques were used to create a comprehensive model capable of addressing both low-profiled attacks and other traffic types. Furthermore, our findings highlight the potential of this approach for enhancing DoS attack detection capabilities and contributing to the development of more robust defense mechanisms.
The power and ubiquity of machine learning demand security measures for protecting sensitive data. Secure multiparty computation (MPC) techniques enable a group of parties to jointly compute a given function while keeping the information private. In this work, we engineer a prototype for privately training support vector machines (SVMs) using MPC techniques. We conduct an extensive study on how different approaches for training SVMs interact with existing state-of-the-art MPC protocols. We identify the least squares (LS) approach as the best suited for privately training. We then optimize fixed-point precision, ensuring accuracy while keeping low running time and communication. The technical details of the optimization involve bounds on the step size of a gradient method to solve a linear system, which might be of independent interest. We further propose and analyse different alternatives to improve the LS approach on an MPC implementation, and we compare their performance. The best improvement yields up to 2x reduction of the running time and communication complexity, without affecting the accuracy of the trained model. In order to illustrate the feasibility of our solution, we securely train SVMs for two realistic tasks.
Network intrusion detection systems based on deep learning are gaining significant traction in cyber security due to their high prediction accuracy and strong adaptability to evolving cyber threats. However, a serious drawback is their vulnerability to evasion attacks that rely on adversarial examples. To provide robustness guarantees for deep neural networks against any possible perturbations, certified defenses against perturbations within a lp-bounded region around the input are being increasingly explored. Unfortunately, unlike existing image domain approaches that concentrate on homogeneous input feature spaces, the progress on certified defense for the network traffic domain, which is characterized by heterogeneous features, has been very limited. To address such a gap, we present the design and practicality of a novel framework, Multi-order Adaptive Randomized Smoothing (MARS), for certifying the robustness of network intrusion detectors based on deep neural networks. Experiments on various network intrusion detection systems show that MARS significantly improves the tightness of robustness certification (12.23\(\%\) increase in l2 certified radius), detection accuracy on evasion attack (7.17\(\%\) improvement on l∞-PGD, 10.11\(\%\) improvement on l1-EAD), and prediction accuracy on natural corruption (16.65\(\%\) enhancement on latency, 18.23\(\%\) enhancement on packet loss) compared to the SOTA method. We have also conducted an extensive analysis of the dimension-wise certified robustness of the network intrusion detector. The results indicate that the dimensional certified radii obtained using MARS reveal the robustness differences across feature dimensions, aligning with the empirical evaluation findings.
The increasing complexity of deep neural networks (DNNs) poses significant resource challenges for edge devices, prompting the development of compression technologies like model quantization. However, while improving model efficiency, quantization can introduce or perpetuate the original model’s bias. Existing debiasing methods for quantized models often incur additional costs. To address this issue, we propose FairQuanti, a novel quantization approach that leverages neuron role contribution to achieve fairness. By distinguishing between biased and normal neurons, FairQuanti employs mixed precision quantization to mitigate model bias during the quantization process. FairQuanti has four key differences from previous studies: (1) Neuron Roles - It formally defines biased and normal neuron roles, establishing a framework for feasible model quantization and bias mitigation; (2) Effectiveness - It introduces a fair quantization strategy that discriminatively quantizes neuron roles, balancing model accuracy and fairness through Bayesian optimization; (3) Generality - It applies to both structured and unstructured data across various quantization bit levels; (4) Robustness - It demonstrates resilience against adaptive attacks. Extensive experiments on five datasets (three structured and two unstructured) using five different models validate FairQuanti’s superior performance against eight baseline methods. Specifically, fairness metrics such as demographic parity (DP) improve by approximately 1.03 times, and the demographic parity ratio (DPR) improves by approximately 1.51 times compared to the baselines, with an average accuracy loss of less than 7.5% at 8-bit quantization. FairQuanti presents a promising solution for deploying fair and efficient deep models on resource-constrained devices and holds potential for application in large language models to reduce size and computational demands while minimizing bias. Our source code is available at https://github.com/Caozq2/FairQuanti.
Despite significant effort put into research and development of defense mechanisms, new malware is continuously developed rapidly, making it still one of the major threats on the Internet. For malware to be successful, it is in the developer's best interest to evade detection as long as possible. One method in achieving this is using Code Injection, where malicious code is injected into another benign process, making it do something it was not intended to do. Automated detection and characterization of Code Injection is difficult. Many injection techniques depend solely on system calls that in isolation look benign and can easily be confused with other background system activity. There is therefore a need for models that can consider the context in which a single system event resides, such that relevant activity can be distinguished easily. In previous work, we conducted the first systematic study on code injection to gain more insights into the different techniques available to malware developers on the Windows platform. This paper extends this work by introducing and formalizing Behavior Nets: A novel, reusable, context-aware modeling language that expresses malicious software behavior in observable events and their general interdependence. This allows for matching on system calls, even if those system calls are typically used in a benign context. We evaluate Behavior Nets and experimentally confirm that introducing event context into behavioral signatures yields better results in characterizing malicious behavior than the state of the art. We conclude with valuable insights on how future malware research based on dynamic analysis should be conducted.
Interdependent systems, with multiple interconnected assets, face escalating cybersecurity threats from external attackers. This article explores security decision-making, operating on complex interdependent systems and proposes a security resource allocation methodology to enhance their proactive security. Using attack graphs, we model vulnerabilities and propose different defense mechanisms integrating different network analysis algorithms, including degree, betweenness, and harmonic centralities, TrustRank, and Katz centrality. We introduce Average Based Node Ranking (ABNR) to average ranks from these methods. The resource allocation methods leverage four different graph-theoretic methods. Each ranking algorithm is combined with these four allocation techniques. Our methods show low sensitivity to simultaneous attacks on interdependent systems. We validate our framework using 11 attack graphs representing real-world systems, measuring security improvements against four well-known allocation algorithms: behavioral decision-making, defense-in-depth, risk-based defense, and min-cut. Our framework outperformed the baselines in most cases, with superior outcomes confirmed by the Friedman statistical test. We show that the main components in our framework have low-time overhead. We also evaluate our framework against multi-stage attacks and cascading failures Our framework enhances security decision-making across different scenarios, including top-1 and all attack paths for different attacks. We release the implementation of our resource allocation methodology to the research community
Presently, cloud computing stands as a dependable choice for enterprises seeking contemporary, adaptable IT solutions capable of managing vast volumes of business data. Its adoption holds the promise of enhancing operational efficiency and productivity. However, cloud computing remains a dynamic and evolving technology landscape, fraught with inherent security challenges. Malevolent actors perpetually scour for novel methodologies to compromise the integrity of data hosted within cloud environments. For instance, data theft, achieved through downloading or encrypting sensitive information, and Distributed Denial of Service (DDoS) assaults targeting cloud infrastructures, pose persistent threats. To address these pressing concerns, the solution outlined in this article advocates for intrusion detection within cloud environments employing a plethora of classification algorithms. To ensure the precision of outcomes, the proposed approach incorporates the meticulous selection of pertinent attributes from the dataset, leveraging the Boruta algorithm. Our research has demonstrated that combining Boruta with classifiers yields impressive results, achieving a recall of 100% with KNN on the CICIDS 2017 dataset and a precision of 100% with Naive Bayes on the CICDDOS 2019 dataset. These results underscore the significant role of feature selection in enhancing detection performance, affirming its importance for achieving optimal results in intrusion detection systems.
False Data Injection Attacks (FDIAs) that target the state estimation pose an immense threat to the security of power grids. Deep Neural Network (DNN)-based methods have shown promising results in detecting such FDIAs. Among the existing state-of-the-art DNN models, time series analysis DNNs have demonstrated superior FDIA detection capability. This article discusses the challenges associated with applying time series analysis DNNs for detecting FDIAs and emphasizes the impact of the attack rate on the detection rate of attacks. We demonstrate that existing time series analysis DNNs are highly vulnerable to FDIAs executed at low attack rates. This article presents various alternative implementations for time series classifiers and time series predictors to improve the FDIA detection rate. A novel method is proposed to train time series classification neural networks to detect FDIAs of any attack rate with high efficiency. Subsequently, an enhanced FDIA detection framework that includes a time series classifier and multiple predictors is presented. Furthermore, an analytical criterion is derived to estimate the FDIA detection rate of time series analysis DNNs under any attack rate. Experimental results obtained on IEEE bus systems using state-of-the-art DNN architectures support the effectiveness of the proposed training method and the proposed framework. The proposed training method significantly improved the detection rate of FDIAs at low attack rates. Up to a 48% improvement in the FDIA detection rate was observed in the proposed framework when compared to the state-of-the-art.
