1393 research outputs found
Sort by
A Novel Approach to Quantum-Resistant Selective Encryption for Agricultural Sensors with Limited Resources
No full textSelective Encryption involves extracting critical features from the data and applying highly secure encryption to those features, while the remaining data is stored using less resource-intensive encryption techniques. Discrete Wavelet Transforms (DWT) provides a means to extract these essential features. Previous works on selective encryption using DWT have explored hardware-specific implementations, such as using a General Purpose GPU (GPGPU). However, in the context of agricultural images captured by edge devices with limited computational capabilities, leveraging a GPGPU would introduce additional hardware requirements and restrict application potential. We present a selective encryption methodology utilizing parallel CPU processing to accelerate calculations, addressing these limitations. Given the advancements in quantum computing, there is a need to ensure the employed encryption methods are secure against potential quantum attacks. We implement NIST-proposed standards: ML-KEM-1024 for key encapsulation and ML-DSA for signature verification, ensuring quantum-resistant security. Our approach provides a security analysis and performance evaluation. We demonstrate significant visual degradation of encrypted data, with mean PSNR of 4.7201 decibel (dB) and SSIM of 0.0003, indicating strong resistance to statistical attacks. Performance improvements range from 21.47% to 52.43% compared to full AES-256 encryption across various file sizes. We discuss optimizations for handling different data sizes and compare our approach\u27s security and performance with existing state-of-the-art methods. This MLPQE method offers a balanced solution for securing agricultural images on resource-constrained edge devices while ensuring long-term data protection against emerging quantum threats
A robust adversarial ensemble with causal (feature interaction) interpretations for image classification
No full textDeep learning-based discriminative classifiers, despite their remarkable success, remain vulnerable to adversarial examples that can mislead model predictions. While adversarial training can enhance robustness, it fails to address the intrinsic vulnerability stemming from the opaque nature of these black-box models. In this paper, we present a deep ensemble model that combines discriminative features with generative models to achieve both high classification accuracy and strong adversarial robustness. Our approach integrates a bottom-level pre-trained discriminative network for feature extraction with a top-level generative classification network that models adversarial input distributions through a deep latent variable model. Using variational Bayes, our model achieves superior robustness against diverse white-box adversarial attacks without requiring adversarial training. Extensive experiments on CIFAR-10 and CIFAR-100 demonstrate our model’s superior adversarial robustness. Through evaluations using counterfactual metrics and feature interaction-based metrics, we establish correlations between model interpretability and adversarial robustness. Our architecture’s generative component is generalizable and can serve as an auxiliary network adaptable to various pre-trained discriminative models. We demonstrate this generalizability through experiments on Tiny-ImageNet with different backbone architectures, indicating the potential applicability of our approach to larger-scale classification datasets
C2PROBER: A Framework to Identify and Label C2 Frameworks
No full textFingerprinting refers to the process of identifying detailed information about a system, a technique employed by both defenders and attackers. Active fingerprinting is crucial for identifying dynamic and short-lived Command and Control (C2) infrastructure. Although passive fingerprinting techniques are widely used in security, active fingerprinting enhances visibility by directly interacting with target systems and servers. Existing detection methods, whether passive monitoring or active probing, identify servers as malicious but rarely attribute them to a specific framework. This study aimed to design and validate an active fingerprinting artifact, C2PROBER, capable of identifying and labeling open-source C2 frameworks through HTTP and TLS probing. The artifact was developed using Python and created a YAML-based rule that defined unique requests to probe, signatures, and confidence-scoring logic to identify and label C2 frameworks. Each probed request triggered a unique response behavior produced through customized HTTP methods, protocol version, URIs, header manipulation, and TLS fingerprint extraction via JARM and X.509 certificate analysis. This research addressed two research questions: 1) Can active fingerprinting techniques identify and label C2 frameworks? 2) Can we effectively parse the HTTP response of the C2 server to obtain unique artifacts (such as response headers, error messages, and patterns) to identify the C2 frameworks Sliver, Empire, and Metasploit? This research methodology followed Wieringa’s Design Science Research (DSR) and validated the artifact using a Single-Case Mechanism Experiment (SCME). The evaluation metrics, including the confusion matrix, precision, recall, and F1 score, demonstrated high detection accuracy, indicating the artifact’s ability to differentiate between C2 infrastructure and benign servers with minimal false positives. The findings revealed that active probing remains a practical approach for identifying and labeling C2 frameworks. This research contributes to a modular, reproducible, and extensible fingerprinting mechanism. Additionally, this study advances fingerprinting from detection to actionable attribution, thus improving threat intelligence and defense
Prioritizing Privacy in Artificial Intelligence: A Practitioner-Grounded, Layered Framework
No full textAs artificial intelligence (AI) systems rapidly assume responsibility for processing large volumes of sensitive personal data, organizations struggle to identify which privacy safeguards deserve immediate, sustained investment. This study develops a practitioner-informed framework that ranks and organizes privacy controls for AI environments, thereby bridging the persistent gap between regulatory mandates and implementation realities. Using grounded theory and thematic analysis of 152 qualitative responses drawn from experts in technology, healthcare, government, defense, and education, we coded interview data in ATLAS.ti to surface three interdependent control layers—Strategic Governance Controls, Operational Controls, and Technical Controls—that map closely to the NIST Privacy Framework and expand the People, Process, Data, and Technology (2PDT) model. Strategic Governance emphasizes executive-level accountability, ethics committees, and privacy-by-design mandates embedded as early as Step Zero of the NIST Risk Management Framework; Operational controls translate these mandates into day-to-day routines such as accounting of disclosures, contractor oversight, privacy training, incident response, and iterative Privacy Impact/Risk Assessments; Technical Controls supply the enabling mechanisms, including data-quality assurance, encryption, anonymization, access control, continuous monitoring, and data minimization-retention practices. Expert rankings assign the highest criticality to data quality and integrity, privacy-enhanced design, and continuous monitoring, underscoring a consensus that reliable data and built-in privacy architectures are foundational to trustworthy AI. The resulting framework delivers actionable guidance for developers, auditors, and policymakers: it clarifies which controls must be prioritized, demonstrates how socio-technical coordination sustains compliance, and offers concrete touchpoints for auditing AI systems against evolving standards such as GDPR, HIPAA, CCPA, and emerging U.S. AI-governance directives. By aligning technical measures with strategic oversight and operational accountability, the study provides a defensible roadmap for privacy-by-design in AI, reduces implementation uncertainty, and strengthens stakeholder trust in data-driven innovation
Leveraging Internet Speed Tests as a Covert Channel for Data Exfiltration
No full textInformation is stolen predominantly from an entity\u27s digital devices and networks through data exfiltration. As defensive technologies advance in detection techniques, threat actors have developed stealthier methods to steal information, including through network-based covert channels. This paper introduces preliminary research on leveraging Ookla\u27s Speedtest traffic as a covert channel for data exfiltration
Automating Mobile Malware Evasion: An Opcode N-Gram Based Approach to Generate Machine Learning Powered Adversarial Examples
Get PDFAs mobile devices continue to become more integrated into daily life, they have become a greater focus area for malicious actors. The Android operating system represents a very enticing target as their products remain dominant, not just in the smartphone market, but also amongst a broad range of smart devices. Because of this, there has been a consistent rise in both the number and sophistication of Android based malware attacks. Anti-virus solutions have, at times, struggled to keep up, as mobile devices are designed with convenience in mind and security solutions tend to be hidden in the background. The lack of user interactions with security solutions limits their full potential. Fortunately, one relatively new innovation is the advent of malware classification AI models, which greatly improves the autonomous capabilities of an anti-virus solution.
Instead of relying on more outdated methods such as signature hashes, which is trivial to circumvent, machine learning can often replicate a more holistic, even human-like, approach to malware analysis. However, underneath the surface, these malware classification models are still dependent on basic logic such as matching opcodes or looking for specific API calls. Because of this certain classification models are still susceptible to be deceived with even just a few small modifications to the original file. This approach can be made even easier by utilizing generative adversarial networks, or GANs, to automatically identify the changes that need to be made and output files that are resistant to malware classification.
Just as security researchers have utilized machine learning to defend against Android malware, malicious actors also have the opportunity to leverage the same technologies to defeat them. This represents a potentially dangerous future where malware authors can quickly deploy stealthy viruses that can not only bypass older antivirus solutions, but some of the most modern ones as well. This paper will outline the context surrounding AI generated malware, propose a potential methodology for creating an Android specific solution in this field, lay out the work that will be necessary to achieve this endeavor, and finally form a timeline in which this work will be completed
Does the Use of Audits Decrease the Infection Rate in a Medical Care Setting?
No full textNosocomial infections are a significant health concern in medical settings. Reports of low compliance rates with hand hygiene standards, guidelines of which are outlined and mandated by the CDC and WHO, are frequent. Factors contributing to nonadherence include lack of knowledge and an unclear understanding of correct techniques (4). Evidence shows that improved hand hygiene can reduce infection rates (1), especially when healthcare providers are included in interventions that aim to improve compliance (4).https://scholar.dsu.edu/research-symposium/1064/thumbnail.jp
SURVEY OF PRAIRIE LAKES FOR MICROPLASTICS IN EASTERN SOUTH DAKOTA
Get PDFMicroplastics are plastics that have been worn down into small fragments that invade aquatic environments and have cultivated themselves into most corners of the world. They are defined as being less than 5 mm in size and categorized as either primary (manufactured small) or secondary (broken down from larger plastics) (Calcutt, Jo, et al., 2018). Microplastics have been found in ocean surface water and deep-sea benthic zones as well as freshwater systems. They can be harmful to organisms, ecosystems, and human health, though the full extent of their impact is not yet known.https://scholar.dsu.edu/research-symposium/1059/thumbnail.jp
Enhancing Crop Yield through Efficient Anomaly Detection Using Transfer Learning and Multispectral Satellite Imagery
No full textThe increasing demand for sustainable agriculture necessitates innovative approaches for monitoring and enhancing crop health. Data-driven methods, combined with advanced machine learning models and remote sensing technologies, present significant potential to bridge the gap between early anomaly detection and timely intervention. Our research explores the development of a robust system integrating state-of-the-art deep learning techniques with transfer learning and multispectral satellite imagery to detect crop anomalies. The proposed system leverages publicly available datasets to identify early symptoms of crop stress—such as yellowing, spotting, and wilting—in key crops, including corn, soybeans, wheat, and sunflowers. Furthermore, the study investigates the influence of environmental factors, such as lighting conditions, weather patterns, and soil characteristics, on detection accuracy. By providing actionable insights for optimizing intervention strategies, this research aims to advance sustainable agricultural practices and improve crop yields. The work also contributes to the broader field of data science by demonstrating the application of sophisticated models in tackling complex agricultural challenges
Impact of Data Snooping on Deep Learning Models for Locating Vulnerabilities in Lifted Code
Get PDFThis study examines the impact of data snooping on neural networks used to detect vulnerabilities in lifted code, and builds on previous research that used word2vec and unidirectional and bidirectional transformer-based embeddings. The research specifically focuses on how model performance is affected when embedding models are trained with datasets, which include samples used for neural network training and validation. The results show that introducing data snooping did not significantly alter model performance, suggesting that data snooping had a minimal impact or that samples randomly dropped as part of the methodology contained hidden features critical to achieving optimal performance. In addition, the findings reinforce the conclusions of previous research, which found that models trained with GPT-2 embeddings consistently outperformed neural networks trained with other embeddings. The fact that this holds even when data snooping is introduced into the embedding model indicates GPT-2\u27s robustness in representing complex code features, even under less-than-ideal conditions