Research / Paper Files

Paper Files

Paper Files are a personal practice of continuous study in cybersecurity that applies the scientific method as a learning tool. Each article is written not for academic publication, but as a deliberate exercise to sharpen the axe — deepening knowledge on specific topics, organizing ideas with rigor, and observing how understanding of certain subjects develops over time. It is the intellectual equivalent of a training journal: each entry records what was studied, how it was analyzed, and what conclusions emerged.

I strongly recommend that security professionals adopt this method. Choose a topic you want to master, research primary sources (RFCs, papers, official documentation), structure your text with introduction, methodology, analysis, and conclusion, and write as if you were explaining to a colleague. The act of writing forces clarity of thought and exposes gaps in understanding that passive reading cannot reveal. Use the articles on this site as examples of format and depth, and build your own collection — over time, you will have a personal repository of refined knowledge that no course or certification can replace.

DAI & ADAI · 2026

DAI · EN

DAI: A Domain-Specific LLM with Model-Excluded Authority, Capability-Disjoint Verification, and Recomputable Receipts

DAI (Decripte Artificial Intelligence) is a proprietary domain-specific LLM within a governed autonomous defensive architecture. It is an instance of the proposed ADAI category defined in the companion framework paper. T…

Read paper →ADAI · EN

Autonomous Defensive AI: An Envelope-and-Evidence Framework for Governed LLMs Acting on Production Infrastructure

ADAI (Autonomous Defensive Artificial Intelligence) is a proposed category of governed autonomous cyber defense. The framework specifies defended perimeters, bounded effects, declared rollback and capability-disjoint ver…

Read paper →

All Detection DFIR AppSec DevSecOps AI Security IAM Cloud Containers IR Crypto Research Phishing Secrets

13 papers

AI Security

Integrated Defense Architecture for AI-Based Systems: Formal Threat Modeling, Adversarial Attack Detection, and Operational Monitoring in Production Environments

TITLE: Integrated Defense Architecture for AI-Based Systems: Formal Threat Modeling, Adversarial Attack Detection, and Operational Monitoring in Production Environments ABSTRACT: The adoption of machine learning-based systems in corporate environments has transformed artificial intelligence models into critical components of operational decision-making. These systems are now used in fraud detection, malware classification, image-assisted medical triage, recommendation systems, and behavioral analytics on digital platforms. However, unlike traditional software, machine learning models are parametric functions trained from data, which introduces additional attack surfaces associated with input manipulation, contamination of the training process, and exploitation of inference interfaces. Recent literature documents multiple classes of attacks against machine learning systems, including adversarial examples, data poisoning, model extraction, and membership inference. At the same time, the integration of these systems into DevOps pipelines and container-based infrastructures introduces additional risks related to the software and model supply chain. This work proposes an integrated defense architecture for artificial intelligence systems deployed in production, combining formal threat modeling, statistical anomaly detection mechanisms, cryptographic integrity verification, and continuous operational monitoring. We formalize the concept of aggregated adversarial risk as a function of conditional probabilities associated with different attack classes under a parameterized adversary model. We then introduce a hybrid detector based on Mahalanobis distance and the Isolation Forest algorithm applied to latent model representations. Complementarily, we propose an integrity verification mechanism that combines cryptographic hashing, digital signatures, and behavioral tests based on output distribution divergence. The architecture is integrated into a Kubernetes-based operational pipeline, with telemetry correlated to techniques described in frameworks such as MITRE ATT&CK and MITRE ATLAS. Experimental evaluations in controlled scenarios indicate a consistent increase in empirical robustness against evasion attacks and improved detection of unauthorized changes in model artifacts. The analysis also shows moderate latency impacts when defense mechanisms are properly dimensioned. The results suggest that approaches combining cryptographic controls, statistical analysis, and operational monitoring offer a pragmatically viable path to improving the security of artificial intelligence systems deployed in production.

2008-07-15Read paper →AI Security

Operational Architecture for Security of Artificial Intelligence Systems: Threat Modeling, Defense in Depth, and Empirical Evaluation in Critical Environments

TITLE: Operational Architecture for Security of Artificial Intelligence Systems: Threat Modeling, Defense in Depth, and Empirical Evaluation in Critical Environments ABSTRACT: The rapid incorporation of artificial intelligence (AI) systems into critical infrastructures has introduced new classes of vulnerabilities that are not fully captured by traditional information security models. Unlike deterministic software, machine learning models depend heavily on training data and statistical distributions, creating attack surfaces associated with data manipulation, exploitation of inference interfaces, and compromise of the data supply chain. This work proposes an operational architecture for protecting AI systems based on defense-in-depth principles, integrating formal threat modeling, statistical mechanisms for detecting data manipulation, and operational controls integrated with security operations centers (SOC). We initially formalize a mathematical model to characterize desirable security properties in supervised learning pipelines, including statistical data integrity and parametric model confidentiality. We then introduce data poisoning detection mechanisms based on Kullback–Leibler divergence and sample influence analysis, combined with adaptive query limiting strategies to mitigate model extraction attacks. The proposed architecture integrates security controls throughout the full AI model lifecycle, including data ingestion, training, packaging, and deployment in orchestrated environments. The experimental implementation uses a contemporary technology stack composed of Kubernetes, PyTorch, MLflow, Suricata, and integration with an enterprise SIEM. We empirically evaluate the approach in a laboratory environment representing critical infrastructure, using public security datasets and controlled adversarial scenarios. The results indicate a significant reduction in the effectiveness of adversarial and model extraction attacks, with moderate operational impact in terms of latency and computational resource consumption. A case study in an anti-fraud detection scenario demonstrates the feasibility of the architecture in real operational contexts. We also discuss implications for AI governance and alignment with emerging normative frameworks, such as the NIST AI Risk Management Framework. The results suggest that statistical mechanisms combined with traditional operational controls constitute an effective strategy for reducing risks associated with the growing adoption of AI in critical environments.

2008-07-08Read paper →AI Security

Operational Architecture for Securing Artificial Intelligence Systems: Threat Modeling, Defense in Depth, and Empirical Evaluation in Critical Environments

TITLE: Operational Architecture for Securing Artificial Intelligence Systems: Threat Modeling, Defense in Depth, and Empirical Evaluation in Critical Environments ABSTRACT: The rapid adoption of artificial intelligence (IA) systems in critical infrastructure has introduced new classes of vulnerabilities that are not fully captured by traditional information security models. In contrast to deterministic software, machine learning models depend significantly on training data and statistical distributions, thereby creating attack surfaces associated with data manipulation, exploitation of inference interfaces, and compromise of the data supply chain. This work proposes an operational architecture for defending IA systems based on defense in depth principles, integrating formal threat modeling, statistical mechanisms for detecting data manipulation, and operational controls integrated with Security Operations Centers (SOC). First, we formulate a mathematical model for characterizing desired security properties in supervised learning pipelines, including statistical integrity of the data and parametric confidentiality of the model. Next, we present detection mechanisms for data poisoning based on Kullback–Leibler divergence and sample influence analysis, combined with adaptive query-rate limiting strategies to mitigate model extraction attacks. The proposed architecture incorporates security controls throughout the entire lifecycle of IA models, including data ingestion, training, packaging, and deployment in orchestrated environments. The experimental implementation uses a contemporary technology stack including Kubernetes, PyTorch, MLflow, Suricata, and integration with an enterprise SIEM. We empirically evaluate the approach in a laboratory environment representing critical infrastructure, using public security datasets and controlled adversary scenarios. The results indicate a significant reduction in the effectiveness of adversarial attacks and model extraction attacks, with moderate operational impact in terms of latency and computational resource consumption. A case study in an antifraud detection scenario demonstrates the applicability of the architecture in real operational contexts. We also discuss implications for IA governance and alignment with emerging normative frameworks, such as the NIST AI Risk Management Framework. The results suggest that statistical mechanisms integrated with traditional operational controls constitute an effective strategy for mitigating risks associated with the growing adoption of IA in critical environments.

2008-07-08Read paper →AI Security

Operational Architecture for the Security of Artificial Intelligence Systems: Threat Modeling, Defense in Depth, and Empirical Evaluation in Critical Environments

TITLE: Operational Architecture for the Security of Artificial Intelligence Systems: Threat Modeling, Defense in Depth, and Empirical Evaluation in Critical Environments ABSTRACT: Operational Architecture for the Security of Artificial Intelligence Systems: Threat Modeling, Defense in Depth, and Empirical Evaluation in Critical Environments RESUMO: Operational Architecture for the Security of Artificial Intelligence Systems: Threat Modeling, Defense in Depth, and Empirical Evaluation in Critical Environments Abstract: The rapid integration of artificial intelligence (AI) systems into critical infrastructure has introduced a new class of vulnerabilities that conventional information security models cannot adequately capture. Unlike deterministic software, machine learning models depend heavily on training data and statistical distributions, creating attack surfaces related to data poisoning, abuse of inference interfaces, and supply chain compromise of data. This study proposes an operational architecture for securing AI systems based on the principle of Defense in Depth, integrating formal threat modeling, statistical mechanisms to detect data poisoning, and operational controls integrated into a Security Operations Center (SOC). First, we formalize a mathematical model that characterizes the desired security properties in supervised learning pipelines, namely the statistical integrity of the data and the parametric confidentiality of the models. Next, we introduce a data poisoning detection mechanism based on Kullback–Leibler divergence and sample influence analysis, combined with an adaptive query rate limiting strategy to mitigate model extraction attacks. The proposed architecture integrates security controls throughout the entire AI model lifecycle, including data ingestion, training, packaging, and deployment in orchestration environments. The implementation uses a modern technology stack composed of Kubernetes, PyTorch, MLflow, Suricata, and integration with an enterprise SIEM. We empirically evaluate this approach in a laboratory environment representative of critical infrastructure, using public security datasets and controlled adversarial scenarios. The results showed a substantial reduction in the effectiveness of adversarial attacks and model extraction attacks, with moderate operational impact in terms of latency and computational resource consumption. A case study in a fraud detection scenario demonstrates the feasibility of the architecture in an operational environment. We also discuss implications for AI governance and alignment with emerging normative frameworks, such as the NIST AI Risk Management Framework. These results suggest that combining statistical mechanisms with conventional operational controls is an effective strategy for reducing the risks associated with the growing adoption of AI in critical environments.

2008-07-08Read paper →AI Security

An Integrated Security Architecture for Artificial Intelligence Systems: Formal Threat Modeling, Adversarial Attack Detection, and Operational Hardening in Corporate Environments

TITLE: An Integrated Security Architecture for Artificial Intelligence Systems: Formal Threat Modeling, Adversarial Attack Detection, and Operational Hardening in Corporate Environments ABSTRACT: The rapid incorporation of Artificial Intelligence (AI) systems into critical corporate infrastructures has introduced a new class of security risks that transcends traditional cybersecurity defense models. Deep learning models, inference services exposed via API, and more recently Large Language Models (LLMs) have become part of operational pipelines responsible for sensitive decisions, including fraud detection, automated classification, and incident response. However, recent research demonstrates that such systems exhibit structural vulnerabilities, including adversarial attacks, model extraction, membership inference, and data poisoning. In addition, the operational ecosystem that supports AI applications remains susceptible to traditional infrastructure threats, expanding the attack surface. In this work, we propose an integrated security architecture for AI systems that combines formal threat modeling, adversarial attack detection mechanisms, and operational hardening practices tailored to corporate environments. First, we develop a unified threat model that integrates the MITRE ATT&CK and MITRE ATLAS taxonomies, enabling the mapping of machine learning–specific attack vectors within established cybersecurity defense frameworks. Next, we formalize the adversarial robustness problem from a min–max optimization perspective and introduce a detection mechanism based on output entropy and local intrinsic dimensionality estimation to identify potentially adversarial inputs at inference time. The proposed architecture also incorporates statistical drift monitoring, data pipeline integrity validation, and integration with SIEM platforms for operational response. The implementation was carried out using PyTorch, Suricata, Elastic Stack, and Kubernetes infrastructure, reflecting widely adopted practices in corporate environments up to 2023. Experimental evaluation was conducted in computer vision and intrusion detection scenarios using the CIFAR‑10 and UNSW‑NB15 datasets. The results indicate substantial improvement in robustness against FGSM and PGD attacks, as well as a significant reduction in mean operational response time in security operations centers (SOC). The findings suggest that isolated approaches to adversarial robustness are insufficient and that AI security must be treated as a systemic problem that integrates mathematical methods, secure software engineering, and continuous infrastructure monitoring.

2008-07-01Read paper →AI Security

Formal Architecture for the Security of Artificial Intelligence Systems: Threat Modeling, Adversarial Verification, and Robust Mitigation with Quantifiable Guarantees

TITLE: Formal Architecture for the Security of Artificial Intelligence Systems: Threat Modeling, Adversarial Verification, and Robust Mitigation with Quantifiable Guarantees ABSTRACT: The growing incorporation of Artificial Intelligence (AI) systems into critical infrastructures, financial services, digital platforms, and industrial environments has introduced a new attack surface whose complexity transcends traditional paradigms of computational security. Deep learning models are susceptible to adversarial attacks, data poisoning, model extraction, and sensitive attribute inference, compromising confidentiality, integrity, and availability. In this work, we propose an integrated formal architecture for AI security that combines mathematical threat modeling, automated adversarial verification, and robust mitigation mechanisms with quantifiable guarantees. Our methodology unifies three components: (i) a formal adversary model parameterized by computational budget and access capability, (ii) an anomaly detection mechanism based on adaptive statistical divergence, and (iii) a certified robustness module using relaxed convex optimization. We implemented a prototype in PyTorch and TensorFlow environments, evaluating it on the CIFAR-10, ImageNet-subset, and UNSW-NB15 datasets. Under PGD attacks with ε = 8/255, our method increased robust accuracy from 32.4% (baseline) to 61.7%, with a standard deviation of 1.8%. In poisoning scenarios with 10% corrupted data, we reduced the impact on generalization error by 47%. We also demonstrated a 63% reduction in the success rate of model extraction attacks. The complexity analysis indicates an average overhead of 18% in inference time, while maintaining operational viability. Our results show that the systematic integration of formal verification and statistically grounded mitigation constitutes a substantial advance in the engineering of AI security, providing measurable guarantees aligned with the NIST AI Risk Management Framework.

2007-08-09Read paper →AI Security

Formal Architecture for the Security of Artificial Intelligence Systems: A Unified Defense Model Against Adversarial, Poisoning, and Model Extraction Attacks

TITLE: Formal Architecture for the Security of Artificial Intelligence Systems: A Unified Defense Model Against Adversarial, Poisoning, and Model Extraction Attacks ABSTRACT: The growing integration of Artificial Intelligence (AI) systems into critical infrastructure, financial services, healthcare platforms, and industrial environments has significantly expanded the cyberattack surface. Machine learning (ML) and deep learning (DL) models introduce specific vulnerabilities, including adversarial attacks, data poisoning, and model extraction, which are not adequately mitigated by traditional information security mechanisms. In this work, we propose a unified formal architecture for AI security, integrating statistical robustness mechanisms, runtime anomaly detection, and cryptographic guarantees. Our methodology combines (i) adaptive adversarial regularization, (ii) probabilistic detection based on distribution divergence, and (iii) integrity proof mechanisms using authenticated hash functions and remote verification. We formalize the problem under a parametric threat model and demonstrate that our approach reduces the success rate of adversarial attacks from 78.4% to 12.7% in white-box scenarios and from 64.2% to 18.9% in black-box scenarios. In poisoning attacks with a contamination rate of 10%, we observed a reduction in the impact on accuracy from 22.5% to 4.3%. The average computational overhead was 11.6% during inference and 18.2% during training. We evaluated the system on three widely used datasets (CIFAR-10, ImageNet-Subset, and UNSW-NB15) and compared it against baseline defenses, including classical adversarial training and mechanisms based on isolated statistical detection. The results demonstrate statistically significant superiority (p < 0.01) in robustness and stability. We conclude that AI security should be treated as a formal discipline with verifiable properties, rather than as an ad hoc extension of traditional security.

2007-08-02Read paper →AI Security

Formally Verifiable Architecture for the Security of Machine Learning Models under Adaptive Attacks: An Integrated Framework of Robustness, Monitoring, and Cryptographic Guarantees

TITLE: Formally Verifiable Architecture for the Security of Machine Learning Models under Adaptive Attacks: An Integrated Framework of Robustness, Monitoring, and Cryptographic Guarantees ABSTRACT: The growing integration of machine learning (ML) models into critical infrastructures has introduced a new attack surface in the cybersecurity ecosystem. Adversarial attacks, data poisoning, model extraction, and attribute inference have demonstrated the fragility of systems based on artificial intelligence, with documented impacts in the financial, healthcare, and defense sectors. In this work, we propose an integrated and formally grounded architecture for the security of ML models under adaptive adversaries, combining certified adversarial robustness, runtime detection of behavioral deviations, and cryptographic guarantees of integrity and confidentiality. Our methodology articulates three main components: (i) robust training with Lipschitz regularization and probabilistic certification; (ii) online monitoring based on statistical modeling of latent distributions and anomaly detection via sequential tests; and (iii) cryptographic encapsulation with remote attestation and confidential computing. We formalize the threat model with capabilities aligned with MITRE ATT&CK (T1565, T1190, T1027) and define measurable security properties. We evaluate the architecture on three datasets (CIFAR-10, UNSW-NB15, and MIMIC-III) and two real intrusion detection scenarios. Compared with adversarial robustness baselines (PGD-training) and isolated detection, our framework reduced the attack success rate of adversarial attacks from 78.4% to 21.7%, detected 93.2% of poisoning attacks with a false positive rate of 3.8%, and maintained an average computational overhead of 11.6%. We demonstrate that the systematic integration of formal and operational mechanisms yields substantial gains in systemic resilience, establishing concrete guidelines for the secure development of AI systems in high-risk environments.

2007-07-26Read paper →AI Security