Autonomous Security Systems: The Rise of AI Security Agents

Privacy-Preserving Machine Learning: Securing Training Data for Effective Security Models

I. The Landscape of Privacy-Preserving Machine Learning for Secure Training Data

A. Defining Privacy-Preserving Machine Learning (PPML)

Privacy-Preserving Machine Learning (PPML) represents a confluence of machine learning (ML) and privacy-enhancing technologies (PETs) designed to enable the training and deployment of ML models without exposing sensitive raw data.¹ In an era characterized by the proliferation of big data and the pervasive integration of artificial intelligence (AI), ML algorithms are increasingly leveraged to extract valuable insights from vast datasets. However, these datasets frequently contain confidential information pertaining to individuals, organizations, or proprietary business operations.³ PPML endeavors to strike a balance between harnessing the analytical power of ML and upholding the critical imperative of data privacy.¹ It achieves this by incorporating a variety of privacy-enhancing strategies that permit multiple input sources to collaboratively train ML models, or allow models to make inferences on new data, without leaking private or sensitive information in its original form. The fundamental goal is to prevent data leakage during the ML lifecycle, thereby safeguarding sensitive information while still enabling the development of effective models.

The context for PPML's emergence is the dual trend of escalating data generation and the increasing sophistication of ML algorithms. As organizations accumulate more data, the potential for privacy breaches, whether accidental or malicious, grows commensurately. Traditional ML paradigms often require centralized access to raw data, creating significant privacy vulnerabilities. PPML offers alternative approaches where data can remain decentralized or be transformed in such a way that meaningful analysis is possible without compromising the underlying sensitive details. Common methods employed within PPML include, but are not limited to, differential privacy, federated learning, and homomorphic encryption.²

B. The Imperative: Why Securing Training Data is Crucial for Security Models

The necessity of securing training data takes on heightened importance in the context of security models. These models, designed for applications such as intrusion detection, fraud prevention, malware analysis, and threat intelligence, are themselves critical components of an organization's defense posture. The data used to train these security models often contains exceptionally sensitive information, including details about system vulnerabilities, documented attack patterns, confidential user behaviors, financial transaction records, or even characteristics of critical infrastructure.³

The compromise or leakage of such training data can have severe repercussions. Attackers could gain invaluable insights into the detection mechanisms of security systems, identify blind spots, or understand how to circumvent existing defenses. This not only weakens the security posture but can also directly facilitate future attacks. Furthermore, the training data for security models may itself contain personally identifiable information (PII) or other regulated data types. Its exposure can lead to significant non-compliance issues with stringent data protection regulations like the General Data Protection Regulation (GDPR) in Europe, the California Consumer Privacy Act (CCPA), and the Health Insurance Portability and Accountability Act (HIPAA) in the United States, among others. These regulations mandate robust data protection measures and can impose substantial fines and legal consequences for data breaches or failures in compliance.³

Beyond regulatory and direct security impacts, the failure to protect sensitive training data erodes trust among users, customers, and other stakeholders.³ If individuals or organizations perceive that their data, especially data used to build protective systems, is not handled securely, their willingness to adopt and interact with AI-driven security solutions will diminish. This can hinder the broader deployment and effectiveness of advanced security measures. The increasing frequency, sophistication, and cost associated with data leaks and cyberattacks, which often target large repositories of data explicitly gathered for training ML models, further underscore this imperative.⁴

Moreover, the ML models themselves, once trained, can become targets. Various adversarial attacks, such as membership inference (determining if a specific individual's data was part of the training set), attribute inference (deducing sensitive attributes of training data), model inversion (reconstructing parts of the training data from the model), and property inference attacks, can potentially extract sensitive information from a trained model, even without direct access to the original training dataset.¹ This vulnerability highlights a unique challenge: the tools designed for security (ML models) can inadvertently become conduits for privacy violations if their training data is not adequately protected throughout the ML lifecycle. This creates a "security model paradox" where the protectors themselves require stringent protection for their developmental inputs. PPML aims to address this by enabling the creation of effective security models while minimizing the exposure of the sensitive data upon which they are built.

The drive for PPML adoption is significantly propelled by the rigorous demands of these privacy regulations. The substantial financial and legal penalties for non-compliance often serve as a more immediate and quantifiable motivator for organizations than the abstract risk of a future security breach. Consequently, industries operating under strict regulatory oversight may exhibit faster adoption rates of PPML techniques, viewing them not just as a security enhancement but as a prerequisite for legal and ethical operation.

C. Core Challenges: The Trilemma of Privacy, Utility, and Performance in PPML

The development and deployment of PPML solutions are fundamentally constrained by a complex interplay between three critical dimensions: privacy, utility, and performance (or efficiency). Achieving an optimal balance across these three aspects—often referred to as the PPML trilemma—constitutes a central challenge in the field.⁴

1. Privacy: This dimension refers to the strength and comprehensiveness of the protection afforded to sensitive information within the training data. Stronger privacy guarantees typically involve more aggressive data transformation, perturbation, or restriction of access. The goal is to minimize the risk of data leakage, re-identification, or inference of sensitive attributes.¹⁰
2. Utility: Utility measures the effectiveness, accuracy, and overall usefulness of the ML model trained using privacy-preserving techniques. For security models, high utility is paramount, translating to high detection rates for threats (e.g., intrusions, fraud, malware), low false positive rates (to avoid alert fatigue and unnecessary interventions), and low false negative rates (to ensure actual threats are not missed).¹⁰ Any degradation in these metrics due to privacy measures can undermine the fundamental purpose of the security model.
3. Performance (Efficiency): This encompasses the computational overhead, latency, resource consumption (e.g., processing power, memory), and communication costs introduced by the PPML techniques.⁴ Many advanced PETs, particularly cryptographic ones, can be resource-intensive, potentially making model training or inference too slow or costly for practical deployment in real-time security applications.

Navigating this trilemma is challenging because these dimensions are often inversely correlated. Enhancements in one area frequently lead to compromises in others. For instance:

• Differential Privacy (DP): Achieving stronger privacy guarantees (a lower privacy budget, ε) typically requires adding more statistical noise to the data or model parameters. This increased noise can directly reduce the model's accuracy and utility, as the model learns from a less precise representation of the underlying data.⁴
• Homomorphic Encryption (HE): While HE allows computations on encrypted data, thereby offering strong privacy, it incurs substantial computational overhead. Complex HE schemes that support a wider range of operations or deeper computations (necessary for complex ML models) are often significantly slower and require more resources than operations on plaintext data.¹² This can render model training or real-time inference impractical.
• Federated Learning (FL): By keeping data decentralized, FL enhances privacy. However, it can introduce significant communication costs due to the repeated exchange of model updates between clients and a central server.⁴ Moreover, data heterogeneity across clients can negatively impact the utility of the globally aggregated model.
• Secure Multi-Party Computation (SMPC): SMPC enables collaborative computation without revealing private inputs, but it often involves high communication and computational costs, particularly for complex functions or a large number of parties.⁴

This interconnectedness means that the selection and implementation of a PPML technique is not a straightforward choice of features but rather a careful calibration process within a multi-dimensional optimization space. The specific mechanisms of each PPML technique directly influence how data is represented or how computational steps are performed, which in turn causally impacts model accuracy (utility) and the resources required (performance). Therefore, practitioners must carefully evaluate these trade-offs in the context of their specific security application, data sensitivity, regulatory requirements, and available computational resources. The ideal PPML solution would offer robust privacy with minimal impact on utility and performance, but current technologies often necessitate compromises.

The following table provides a high-level comparative overview of the core PPML techniques that will be discussed in detail in subsequent sections.

Table 1: Comparative Overview of Core PPML Techniques

Technique	Brief Mechanism	Primary Privacy Benefit	Key Utility Impact	Major Performance Overhead	General Scalability	Typical Security Model Use Cases
Differential Privacy (DP)	Adds calibrated statistical noise to data, queries, or model parameters to mask individual contributions.86	Provable guarantee against inferring individual data from output; protects against membership/attribute inference.21	Potential reduction in model accuracy, especially for minority classes or with strong privacy (low ε).4	Moderate increase in training time for DP-SGD; noise calibration can be complex.8	Generally good, especially central DP; local DP can be more challenging.10	Anomaly detection, user behavior analysis, training models on sensitive logs where individual activity needs protection.23
Homomorphic Encryption (HE)	Allows computations (e.g., addition, multiplication) directly on encrypted data without decryption.12	Data remains encrypted during processing by untrusted parties (e.g., cloud MLaaS).13	Ideally preserves accuracy, but approximation in some schemes or noise accumulation can affect it; primary impact via performance.11	Very high computational overhead, especially for FHE and complex operations; significant latency.12	Challenging for deep/complex models and large datasets due to overhead.17	Secure model inference in untrusted environments, collaborative analytics on encrypted financial/health data for fraud/threat detection.613
Federated Learning (FL) / Personalized FL (PFL)	Trains a shared global model by aggregating updates from decentralized clients, keeping raw data local. PFL tailors models to clients.2	Raw data does not leave local devices/silos, reducing exposure risk.25	Global model utility can be affected by data heterogeneity (non-IID). PFL aims to improve local utility. Updates can still leak info.10	High communication costs for model updates; coordination overhead.4	Scalable to many clients, but heterogeneity and communication are bottlenecks.28	Collaborative intrusion detection, malware analysis across organizations, IoT device security.29
Secure Multi-Party Computation (SMPC)	Enables multiple parties to jointly compute a function on their private inputs without revealing inputs to each other.32	Allows collaborative analysis without a trusted third party or sharing raw data among participants.33	Accuracy depends on arithmetic precision (e.g., fixed-point); utility primarily impacted by performance constraints.4	High communication and computational overhead, especially for complex functions or many parties.4	Limited by network latency and complexity of protocols for large-scale, interactive computations.33	Joint threat intelligence sharing, collaborative fraud detection, secure statistical analysis on distributed sensitive datasets.33
Advanced Data Anonymization	Modifies data (e.g., generalization, suppression, shuffling) to prevent re-identification of individuals.14	Reduces risk of linking data to individuals, helps meet compliance if data becomes non-personal.5	Can significantly reduce data utility and model accuracy due to information loss, especially with aggressive methods.14	Varies: simple methods (k-anonymity) have low overhead; dynamic/advanced methods can be more complex.14	Depends on technique; managing anonymization across large, evolving datasets can be complex.	Pre-processing sensitive log data, de-identifying datasets for research or sharing where some utility loss is acceptable.5
Synthetic Data Generation (SDG)	Creates artificial data that mimics statistical properties of real data without containing real individual records.39	Enables model training and data sharing without exposing real sensitive data; can be combined with DP.40	Utility depends on fidelity of synthetic data to real data; can be excellent if generator is good, but may miss nuances or amplify biases.39	Generators (GANs, VAEs) can be computationally intensive to train; generation itself might be faster.42	Depends on generation method; generating diverse, high-dimensional data at scale is challenging.44	Generating rare event data for IDS/fraud, creating privacy-safe datasets for security research, training models where real data is too sensitive/scarce.42

II. Foundational Cryptographic Techniques in PPML (Click to Expand)

A. Homomorphic Encryption (HE): Performing Computations on Encrypted Data

Homomorphic Encryption is a transformative cryptographic primitive that allows computations to be performed directly on encrypted data (ciphertexts) without the need for prior decryption.²¹² The defining characteristic of HE is that if one encrypts some data and then performs homomorphic operations on the ciphertext, the result, when decrypted, is identical to what would have been obtained by performing the equivalent operations on the original plaintext data.¹² This capability is rooted in the algebraic concept of homomorphism, where a function preserves the structure of operations between two algebraic systems.¹⁷ HE is particularly valuable in scenarios where data must be processed by untrusted parties, such as cloud-based Machine Learning as a Service (MLaaS) platforms, as it ensures that the service provider never sees the sensitive raw data.¹⁷

1. Underlying Mechanisms

The journey of homomorphic encryption began with the conceptualization by Rivest, Shamir, and Adleman in 1978, shortly after their invention of the RSA algorithm.¹² However, realizing a practical scheme that could perform arbitrary computations proved elusive for decades. A major breakthrough came in 2009 when Craig Gentry introduced the first plausible construction for Fully Homomorphic Encryption (FHE).¹²

HE schemes can be broadly categorized based on the types and number of operations they support on ciphertexts¹²:

• Partially Homomorphic Encryption (PHE): These schemes support an unlimited number of a single type of operation, either addition or multiplication, but not both. Examples include RSA and ElGamal (multiplicatively homomorphic) and Paillier (additively homomorphic). PHE schemes are generally more efficient than more complex HE variants.¹²
• Somewhat Homomorphic Encryption (SHE): SHE schemes can perform a limited number of both addition and multiplication operations on ciphertexts. The limitation arises from "noise" that is inherent in most HE constructions. Each operation, particularly multiplication, increases this noise, and if it grows too large, the ciphertext becomes undecryptable. Examples include the BGN (Boneh-Goh-Nissim) and CKKS (Cheon-Kim-Kim-Song) schemes.¹² SHE is often employed in practice to avoid the high computational cost associated with FHE's noise management techniques.¹⁷
• Fully Homomorphic Encryption (FHE): FHE schemes represent the most powerful form of HE, allowing for an arbitrary number of both addition and multiplication operations on ciphertexts, thus enabling the computation of any function.¹² The key innovation that makes FHE possible is a technique called "bootstrapping".¹² Bootstrapping essentially takes a noisy ciphertext and homomorphically decrypts it using an encrypted version of the secret key, resulting in a "fresh" ciphertext with reduced noise, thereby allowing further computations. While theoretically powerful, bootstrapping is a very computationally expensive operation and a major performance bottleneck for FHE schemes.¹² Notable FHE schemes include those based on Gentry's original work, as well as schemes like BGV (Brakerski-Gentry-Vaikuntanathan) and BFV (Brakerski/Fan-Vercauteren).¹²

A critical aspect of practical HE schemes is noise management. Most lattice-based HE schemes, which are currently the most promising, involve adding a small amount of noise during encryption. This noise grows with each homomorphic operation. If the noise exceeds a certain threshold, decryption fails. Multiplications tend to increase noise much more significantly than additions.¹⁷

To improve the efficiency of HE, various techniques are employed. Batching allows multiple plaintext values to be packed into a single ciphertext, enabling Single Instruction, Multiple Data (SIMD) style operations where one homomorphic operation is applied to all packed values simultaneously.¹⁷ The Number Theoretic Transform (NTT) is an algorithm used to speed up polynomial multiplication, which is a fundamental operation in many lattice-based HE schemes.¹⁷

2. Impact on Model Performance (Accuracy, Utility)

Ideally, homomorphic encryption should not degrade the accuracy of machine learning models... However, the reality is more nuanced. The utility of HE-trained models can be affected in several ways:

• Approximation Errors: Some HE schemes, like CKKS, are designed for approximate arithmetic on real numbers... introduces small approximation errors that can accumulate.¹² Experiments on Parkinson's and Heart Disease datasets using an HE-based XGBoost inference showed acceptable relative errors of 0-3%.¹⁷
• Computational Noise: The inherent noise in HE schemes, if not perfectly managed... can subtly alter computation results.¹³
• Performance Limitations Impacting Utility: If an HE-protected model is too slow for practical use... its utility is diminished.¹²
• Model Adaptations: Activation functions like ReLU... may need to be approximated by polynomials, which can affect model behavior and accuracy.¹²

Research is actively exploring ways to mitigate these impacts. Frameworks like TT-TFHE aim to optimize FHE for tabular and image data... The primary challenge remains maintaining both accuracy and performance.¹²¹⁷

3. Computational Overhead, Scalability, and Implementation Complexity

The primary drawback of homomorphic encryption is its significant computational overhead.¹²

• Overhead: Operations on ciphertexts are orders of magnitude slower... FHE computations can be around 10,000 times slower.¹² Ciphertexts are also typically much larger than plaintexts.¹³
• Scalability: Challenging for very deep or wide ML models.¹⁷ Application to RNNs is more difficult.¹²
• Implementation Complexity: Notoriously difficult to use correctly and efficiently.¹² Libraries like Microsoft SEAL, IBM HELib, Google's TF-Encrypted, PALISADE (now OpenFHE), and Pyfhel aim to abstract complexity.¹³

4. Current Difficulties and Future Directions

Current Difficulties:

• Performance: Slow speed remains the biggest bottleneck.¹²
• Application to Complex Models: RNNs and Transformers are challenging.¹²
• Usability and Tooling: Lack of user-friendly tools.¹²
• Noise Management: Cost of bootstrapping or limitations of SHE.¹⁷
• Model Utility: Ensuring high accuracy under HE constraints.¹⁷
• Security Concerns: Potential leakage through query patterns, timing analysis.¹⁷
• Interpretability: Encrypted logic can compromise interpretability.¹⁷
• Library Evolution: Reliance on rapidly evolving libraries.¹⁷

Future Directions:

• Efficiency Improvements: Hardware acceleration (GPUs, FPGAs, ASICs) and software optimizations.¹²
• HE-Friendly Model Architectures: Designing ML models inherently compatible with HE constraints.¹²
• Improved Usability: Higher-level abstraction libraries, standardized APIs.¹²
• Hybrid Approaches: Combining HE with other PETs like DP, SMPC, or FL.¹³
• Broader Application Support: Extending HE capabilities for more operations and data types.¹²

B. Secure Multi-Party Computation (SMPC): Collaborative Privacy

Secure Multi-Party Computation (SMPC)... enables multiple parties to jointly compute a function over their respective private inputs in such a way that these inputs remain confidential.¹⁵ SMPC is designed to protect the privacy of participants' inputs from the other participants involved in the computation itself.³²

1. Underlying Mechanisms

Key properties of SMPC include³³: Input Privacy, Computational Integrity, Adversary Resistance (Semi-Honest or Malicious), Decentralization, and Fairness.

Core techniques underpinning SMPC protocols include²⁰: Garbled Circuits (Yao's Protocol), Secret Sharing, and Oblivious Transfer (OT).

2. Impact on Model Performance (Accuracy, Utility)

SMPC techniques allow for training standard ML models on distributed private data. Impact on accuracy is generally due to cryptographic constraints (e.g., integer arithmetic over finite fields) rather than statistical noise. This can introduce quantization errors if ML models use floating-point numbers.⁴

3. Computational Overhead, Scalability, and Implementation Complexity

SMPC protocols are known for significant overheads⁴: Computational Overhead, Communication Overhead (often dominant), Scalability limitations (network latency, number of parties), and Implementation Complexity.

4. Current Difficulties and Future Directions

Current Difficulties: High Overheads, Network Latency, Fixed-Point Arithmetic Constraints, Key Management, Security Against Advanced Threats, Usability and Integration.⁴³³

Future Directions: Hybrid HE-SMPC Models, Performance Optimization (Algorithmic Improvements, Hardware Acceleration), Scalability Enhancements, Post-Quantum SMPC, Improved Usability.³³

III. Decentralized Learning and Statistical Privacy Approaches (Click to Expand)

A. Federated Learning (FL) and Personalized Federated Learning (PFL): Training on Decentralized Data

Federated Learning has emerged as a paradigm-shifting approach... enabling collaborative training of models across multiple decentralized devices or data silos... without requiring the raw data to leave its original location.²

1. Underlying Mechanisms

The typical FL process involves a central server and participating clients⁴: Initialization, Local Training, Update Communication, Aggregation (e.g., FedAvg⁵²), Iteration.

FL can be categorized as Horizontal Federated Learning (HFL)²⁸ or Vertical Federated Learning (VFL).²⁸ Personalized Federated Learning (PFL) addresses data heterogeneity (non-IID data).²⁷ Techniques like Secure Aggregation (SA) further enhance privacy.⁵²

2. Impact on Model Performance (Accuracy, Utility)

FL aims for performance comparable to centralized training. Data Heterogeneity (Non-IID Data) is a major challenge.¹⁰ Studies show promising results for FL in security applications, e.g., financial fraud detection (99.95% accuracy with XFL⁵⁶), malware detection (99.80% with FL+CKKS²⁹).

3. Communication Costs, Scalability, and Implementation Complexity

Challenges include Communication Costs, Scalability²⁸, and Implementation Complexity. Frameworks like TensorFlow Federated (TFF), PySyft, FATE, and Flower aim to simplify development.⁵⁴

4. Privacy Considerations and Enhancements (DP, HE, SMPC in FL)

Model updates can leak information.⁴ Common enhancements: FL with Differential Privacy (FL+DP), FL with Homomorphic Encryption (FL+HE), FL with Secure Multi-Party Computation (FL+SMPC).⁴

B. Differential Privacy (DP): Provable Privacy Guarantees

Differential Privacy (DP) has emerged as a gold standard... offering rigorous, mathematically provable guarantees about the privacy of individuals within a dataset.²

1. Underlying Mechanisms

DP introduces calibrated statistical noise. Key concepts: Privacy Budget (ε, epsilon)⁸, (ε,δ)-Differential Privacy, Sensitivity²², Noise Mechanisms (Laplace, Gaussian⁸, Exponential, Randomized Response⁴), Differentially Private Stochastic Gradient Descent (DP-SGD)⁸, Local vs. Central DP.⁶

2. Impact on Model Performance (Accuracy, Utility, especially for Security Models)

Inherent trade-off between privacy and utility.⁴ Impact on IDS (e.g., SECIoHT-FL achieved 95.48% accuracy with ε=0.34²³). DP-SGD can disproportionately impact accuracy for underrepresented subgroups ("the poor become poorer" effect¹⁸¹⁹).

3. Computational Overhead, Scalability, and Implementation Complexity

Noise addition itself is not a major bottleneck.⁴ DP-SGD can introduce costs. Scalability is generally good. Implementation requires careful consideration of sensitivity analysis, privacy budget management⁶³, and noise calibration. Libraries: Google's DP Library, Opacus, Diffprivlib, TensorFlow Privacy, OpenDP.⁶⁹

4. Challenges and Limitations

Privacy-Utility Trade-off⁴, Disparate Impact on Fairness and Accuracy¹⁸, Choosing Epsilon (ε)¹⁸, Data Destruction vs. Obscuring¹⁷, Limited Protection Against All Threats, Interpretability, Composition.¹⁹

Works Cited (Click to Expand/Collapse)

1. dualitytech.com, accessed May 22, 2025, https://dualitytech.com/blog/privacy-preserving-machine-learning/#:~:text=Privacy%20Preserving%20Machine%20Learning%20(PPML)%20is%20a%20technique%20to%20prevent,exposing%20private%20or%20sensitive%20data. ↩
2. What is Privacy-Preserving Machine Learning? - CookieYes, accessed May 22, 2025, https://www.cookieyes.com/knowledge-base/privacy-tech/what-is-ppml/ ↩
3. Privacy-Preserving Machine Learning Techniques, Challenges And Research Directions - IRJET, accessed May 22, 2025, https://www.irjet.net/archives/V11/i3/IRJET-V111360.pdf ↩
4. rjpn.org, accessed May 22, 2025, https://rjpn.org/ijcspub/papers/IJCSP22A1216.pdf ↩
5. Proactive cybersecurity with Gradiant: AnonymiX and Honeypots ..., accessed May 22, 2025, https://gradiant.org/en/blog/proactive-cybersecurity-anonymix-and-honeypots/ ↩
6. Privacy-Preserving Machine Learning for loT-Integrated Smart Grids ..., accessed May 22, 2025, https://www.mdpi.com/1996-1073/18/10/2515 ↩
7. Which of the Following is a Concern in Privacy Preserving Machine ..., accessed May 22, 2025, https://www.byteplus.com/en/topic/452822 ↩
8. arxiv.org, accessed May 22, 2025, https://arxiv.org/abs/2409.01329 ↩
9. State-of-the-Art Approaches to Enhancing Privacy Preservation of Machine Learning Datasets - arXiv, accessed May 22, 2025, https://arxiv.org/pdf/2404.16847 ↩
10. arxiv.org, accessed May 22, 2025, https://arxiv.org/abs/2108.04417 ↩
11. (PDF) Privacy-Preserving Machine Learning: Methods, Challenges and Directions, accessed May 22, 2025, https://www.researchgate.net/publication/353819224_Privacy-Preserving_Machine_Learning_Methods_Challenges_and_Directions ↩
12. (PDF) Fully homomorphic encryption in PPMLAn review, accessed May 22, 2025, https://www.researchgate.net/publication/382576362_Fully_homomorphic_encryption_in_PPMLAn_review ↩
13. (PDF) Balancing Performance and Privacy: The Impact of ..., accessed May 22, 2025, https://www.researchgate.net/publication/390280977_Balancing_Performance_and_Privacy_The_Impact_of_Homomorphic_Encryption_on_AIML_Model_Efficiency ↩
14. dl.gi.de, accessed May 22, 2025, https://dl.gi.de/bitstreams/3099ae09-9bc3-4ece-b719-acb7b770cb9e/download ↩
15. (PDF) Privacy-Preserving Machine Learning: Securing Data in Al Systems - ResearchGate, accessed May 22, 2025, https://www.researchgate.net/publication/380711820_Privacy-Preserving_Machine_Learning_Securing_Data_in_Al_Systems ↩
16. Empirical Analysis of Privacy-Fairness-Accuracy Trade-offs in Federated Learning: A Step Towards Responsible AI - arXiv, accessed May 22, 2025, https://arxiv.org/html/2503.16233v1 ↩
17. Privacy-Preserving Machine Learning (PPML) Inference for Clinically Actionable Models - Scientia, accessed May 22, 2025, https://scientiasalut.gencat.cat/bitstream/handle/11351/12938/privacy-preserving-machine-learning-ppml-inference-clinically-actionable-models-2025.pdf?sequence=1&isAllowed=y ↩
18. Privacy-Preserving Machine Learning | Coding Crossroads, accessed May 22, 2025, https://www.econai.tech/?page_id=246 ↩
19. www.cs.cornell.edu, accessed May 22, 2025, https://www.cs.cornell.edu/~shmat/shmat_neurips19.pdf ↩
20. Privacy-Preserving AI: Techniques & Frameworks - Dialzara, accessed May 22, 2025, https://dialzara.com/blog/privacy-preserving-ai-techniques-and-frameworks/ ↩
21. Protecting Trained Models in Privacy-Preserving Federated Learning | NIST, accessed May 22, 2025, https://www.nist.gov/blogs/cybersecurity-insights/protecting-trained-models-privacy-preserving-federated-learning ↩
22. arXiv:2309.16398v1 [cs.LG] 28 Sep 2023, accessed May 22, 2025, https://arxiv.org/pdf/2309.16398 ↩
23. Privacy-Preserving Federated Learning-Based Intrusion Detection ..., accessed May 22, 2025, https://www.mdpi.com/2079-9292/14/1/67 ↩
24. Security and Efficiency Tradeoffs in Machine Learning for Cloud Platforms - DiVA portal, accessed May 22, 2025, http://www.diva-portal.org/smash/get/diva2:1919492/FULLTEXT01.pdf ↩
25. Privacy-Enhanced AI for loT: Federated Learning in Security Applications - PhilArchive, accessed May 22, 2025, https://philarchive.org/archive/ROHPAF-2 ↩
26. Federated Learning for Cloud and Edge Security: A Systematic Review of Challenges and AI Opportunities - MDPI, accessed May 22, 2025, https://www.mdpi.com/2079-9292/14/5/1019 ↩
27. Privacy Preserving Machine Learning With Federated Personalized ..., accessed May 22, 2025, https://www.computer.org/csdl/journal/oj/2024/01/10691662/20vjuSYD5tu ↩
28. Scalability Challenges in Privacy-Preserving Federated Learning ..., accessed May 22, 2025, https://www.nist.gov/blogs/cybersecurity-insights/scalability-challenges-privacy-preserving-federated-learning ↩
29. (PDF) A Comparative Study of Privacy-Preserving Techniques in ..., accessed May 22, 2025, https://www.researchgate.net/publication/389965680_A_Comparative_Study_of_Privacy-Preserving_Techniques_in_Federated_Learning_A_Performance_and_Security_Analysis ↩
30. International Journal of Multidisciplinary - PhilArchive, accessed May 22, 2025, https://philarchive.org/archive/ADIFLF ↩
31. (PDF) EEFED: Personalized Federated Learning of ..., ResearchGate, accessed May 22, 2025, https://www.researchgate.net/publication/364576645_EEFED_Personalized_federated_learning_of_ExecutionEvaluation_dual_network_for_CPS_intrusion_detection ↩
32. Secure multi-party computation – Wikipedia, accessed May 22, 2025, https://en.wikipedia.org/wiki/Secure_multi-party_computation ↩
33. www.ijraset.com, accessed May 22, 2025, https://www.ijraset.com/best-journal/privacypreserving-technologies-homomorphic-encryption-and-secure-multiparty-computation ↩
34. (PDF) Secure Multi-Party Computation for Machine Learning: A Survey, accessed May 22, 2025, https://www.researchgate.net/publication/379843467_Secure_Multi-Party_Computation_for_Machine_Learning_A_Survey ↩
35. Privacy-Preserving Collaboration Using Cryptography - Digital.gov, accessed May 22, 2025, https://digital.gov/resources/privacy-preserving-collaboration-using-cryptography ↩
36. Privacy Preserving Machine Learning (PPML) is Essential for AI Development - Knowledge, accessed May 22, 2025, https://knowledge.digitaledge.net/compliance/privacy-preserving-machine-learning-ppml-is-essential-for-ai-development/ ↩
37. Data Obfuscation Through Latent Space Projection for Privacy ..., accessed May 22, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC11922095/ ↩
38. Privacy-Preserving Data Mining Techniques for Sensitive Data Analysis - Mathematical Research Institute Journals, accessed May 22, 2025, https://journals.mriindia.com/index.php/ijaece/article/download/127/114 ↩
39. Synthetic data generation: a privacy-preserving approach ... - Frontiers, accessed May 22, 2025, https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2025.1563991/full ↩
40. Synthetic data generation: a privacy-preserving approach to accelerate rare disease research - PMC, accessed May 22, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC11958975/ ↩
41. GDPR and CCPA: Understanding Synthetic Data, Privacy ... - Gretel.ai, accessed May 22, 2025, https://gretel.ai/gdpr-and-ccpa ↩
42. Synthetic Network Traffic Data Generation: A Comparative Study – arXiv, accessed May 22, 2025, https://arxiv.org/html/2410.16326v2 ↩
43. (PDF) Variational autoencoders (vaes) for synthetic data generation, accessed May 22, 2025, https://www.researchgate.net/publication/385592671_Variational_autoencoders_vaes_for_synthetic_data_generation ↩
44. Scaling Laws of Synthetic Data for Language Models - arXiv, accessed May 22, 2025, https://arxiv.org/html/2503.19551v2 ↩
45. Advanced Techniques for Generating Synthetic Test Data - Tonic.ai, accessed May 22, 2025, https://www.tonic.ai/guides/advanced-techniques-synthetic-test-data ↩
46. jjournals.ju.edu.jo, accessed May 22, 2025, https://jjournals.ju.edu.jo/index.php/JMJ/article/download/2712/836 ↩
47. Privacy-preserving Machine Learning Techniques - Eurecom, accessed May 22, 2025, https://www.eurecom.fr/publication/6641/download/sec-publi-6641.pdf ↩
48. Index - Pyfhel 3.4.2 documentation, accessed May 22, 2025, https://pyfhel.readthedocs.io/ ↩
49. Pyfhel - PyPI, accessed May 22, 2025, https://pypi.org/project/Pyfhel/3.1.1/ ↩
50. MPC Library - CoinFabrik, accessed May 22, 2025, https://www.coinfabrik.com/products/mpc-multi-party-computation-library/ ↩
51. PPML Introduction — BigDL v2.3.0 documentation - Read the Docs, accessed May 22, 2025, https://bigdl.readthedocs.io/en/v2.3.0/doc/PPML/Overview/intro.html ↩
52. www.arxiv.org, accessed May 22, 2025, https://www.arxiv.org/pdf/2505.01788 ↩
53. Federated personalized learning with Differential Privacy algorithm ..., accessed May 22, 2025, https://www.researchgate.net/figure/Federated-personalized-learning-with-Differential-Privacy-algorithm_fig3_384302421 ↩
54. What frameworks are available for federated learning? – Milvus, accessed May 22, 2025, https://milvus.io/ai-quick-reference/what-frameworks-are-available-for-federated-learning ↩
55. arxiv.org, accessed May 22, 2025, https://arxiv.org/abs/2503.16233 ↩
56. Secure and Transparent Banking: Explainable AI-Driven Federated ..., accessed May 22, 2025, https://www.mdpi.com/1911-8074/18/4/179 ↩
57. How Privacy Enhancing Technologies (PETs) Can Help Orgs, accessed May 22, 2025, https://techgdpr.com/blog/discover-how-privacy-enhancing-technologies-pets-help-organizations-achieve-gdpr-compliance-by-safeguarding-personal-data-reducing-risks-and-enhancing-confidentiality-through-encryption-anonymiza/ ↩
58. arxiv.org, accessed May 22, 2025, https://arxiv.org/pdf/2501.15038 ↩
59. arxiv.org, accessed May 22, 2025, https://arxiv.org/abs/2503.09192 ↩
60. [2410.21547] Personalized Federated Learning with Mixture of Models for Adaptive Prediction and Model Fine-Tuning - arXiv, accessed May 22, 2025, https://arxiv.org/abs/2410.21547 ↩
61. What tools are available for simulating federated learning? – Milvus, accessed May 22, 2025, https://milvus.io/ai-quick-reference/what-tools-are-available-for-simulating-federated-learning ↩
62. Trustworthy Federated Learning: Privacy, Security, and Beyond - arXiv, accessed May 22, 2025, https://arxiv.org/html/2411.01583v1 ↩
63. Recent Advances of Differential Privacy in Centralized Deep Learning: A Systematic Survey, accessed May 22, 2025, https://www.researchgate.net/publication/388874871_Recent_Advances_of_Differential_Privacy_in_Centralized_Deep_Learning_A_Systematic_Survey ↩
64. bpb-us-e1.wpmucdn.com, accessed May 22, 2025, https://bpb-us-e1.wpmucdn.com/sites.gatech.edu/dist/c/679/files/2018/09/GDPRDiffPrivacy.pdf?bid=679 ↩
65. Wildest Dreams: Reproducible Research in Privacy-preserving Neural Network Training - arXiv, accessed May 22, 2025, https://arxiv.org/pdf/2403.03592 ↩
66. arXiv:2409.01329v2 [cs.LG] 11 Dec 2024, accessed May 22, 2025, https://arxiv.org/pdf/2409.01329 ↩
67. arxiv.org, accessed May 22, 2025, https://arxiv.org/abs/2505.05843 ↩
68. The Impact of Differential Privacy on Recommendation Accuracy and Popularity Bias, accessed May 22, 2025, https://www.researchgate.net/publication/378985201_The_Impact_of_Differential_Privacy_on_Recommendation_Accuracy_and_Popularity_Bias ↩
69. A Survey of Differential Privacy Frameworks - OpenMined, accessed May 22, 2025, https://openmined.org/blog/a-survey-of-differential-privacy-frameworks/ ↩
70. You Still See Me: How Data Protection Supports the Architecture of ML Surveillance - arXiv, accessed May 22, 2025, https://arxiv.org/html/2402.06609v1 ↩
71. GDPR and unstructured data: is anonymization possible? - Oxford Academic, accessed May 22, 2025, https://academic.oup.com/idpl/article/12/3/184/6552802 ↩
72. Addressing Data Scarcity with Synthetic Data: A Secure and GDPR-compliant Cloud-Based Platform - ResearchGate, accessed May 22, 2025, https://www.researchgate.net/publication/391317294_Addressing_Data_Scarcity_with_Synthetic_Data_A_Secure_and_GDPR-compliant_Cloud-Based_Platform ↩
73. AI Model Training with Synthetic Data: Enabling Privacy-Compliant and Explainable AI Solutions - ResearchGate, accessed May 22, 2025, https://www.researchgate.net/publication/390222058_AI_Model_Training_with_Synthetic_Data_Enabling_Privacy-Compliant_and_Explainable_AI_Solutions ↩
74. (PDF) SYNTHETIC DATA GENERATION FOR QUALITY ASSURANCE IN LARGE-SCALE AI MODELS - ResearchGate, accessed May 22, 2025, https://www.researchgate.net/publication/390411572_SYNTHETIC_DATA_GENERATION_FOR_QUALITY_ASSURANCE_IN_LARGE-_SCALE_AI_MODELS ↩
75. What Is Synthetic Data? Uses, Benefits, Challenges and More - Cohere, accessed May 22, 2025, https://cohere.com/blog/what-is-synthetic-data ↩
76. arxiv.org, accessed May 22, 2025, https://arxiv.org/abs/2410.16326 ↩
77. Leveraging VAE-Derived Latent Spaces for Enhanced Malware Detection with Machine Learning Classifiers – arXiv, accessed May 22, 2025, https://arxiv.org/html/2503.20803v1 (also see pdf version) ↩
78. Leveraging VAE-Derived Latent Spaces for Enhanced Malware Detection with Machine Learning Classifiers – arXiv, accessed May 22, 2025, https://arxiv.org/pdf/2503.20803 ↩
79. Generative adversarial networks (GANs) for synthetic dataset generation with binary classes | Data Science Campus, accessed May 22, 2025, https://datasciencecampus.ons.gov.uk/projects/generative-adversarial-networks-gans-for-synthetic-dataset-generation-with-binary-classes/ ↩
80. Bridging AI and Privacy: Solutions for High-Dimensional Data and Foundation Models - DiVA portal, accessed May 22, 2025, https://diva-portal.org/smash/get/diva2:1955416/FULLTEXT01.pdf ↩
81. Machine Learning for Cyber Threat Detection A Technical Guide for Students & Professionals - ResearchGate, accessed May 22, 2025, https://www.researchgate.net/publication/391369917_Machine_Learning_for_Cyber_Threat_Detection_A_Technical_Guide_for_Students_Professionals ↩
82. [2307.07023] A Controlled Experiment on the Impact of Intrusion Detection False Alarm Rate on Analyst Performance - arXiv, accessed May 22, 2025, https://arxiv.org/abs/2307.07023 ↩
83. Federated Learning-Based Credit Card Fraud Detection: A Comparative Analysis of Advanced Machine Learning Models - ITM Web of Conferences, accessed May 22, 2025, https://www.itm-conferences.org/articles/itmconf/pdf/2025/01/itmconf_dai2024_01022.pdf ↩
84. Malware Detection using ML/DL - PhilArchive, accessed May 22, 2025, https://philarchive.org/archive/PRAMDU ↩
85. GDPR Article 25 | Imperva, accessed May 22, 2025, https://www.imperva.com/learn/data-security/gdpr-article-25/ ↩
86. Guidelines 4/2019 on Article 25 Data Protection by Design and by Default, accessed May 22, 2025, https://www.edpb.europa.eu/sites/default/files/files/file1/edpb_guidelines_201904_dataprotection_by_design_and_by_default_v2.0_en.pdf ↩
87. Art. 25 GDPR – Data protection by design and by default, accessed May 22, 2025, https://gdpr-info.eu/art-25-gdpr/ ↩
88. The Effectiveness of Homomorphic Encryption in Protecting Data Privacy - ResearchGate, accessed May 22, 2025, https://www.researchgate.net/publication/385818007_The_Effectiveness_of_Homomorphic_Encryption_in_Protecting_Data_Privacy ↩
89. Art. 32 GDPR – Security of processing - General Data Protection Regulation (GDPR), accessed May 22, 2025, https://gdpr-info.eu/art-32-gdpr/ ↩
90. GDPR Article 32 | Imperva, accessed May 22, 2025, https://www.imperva.com/learn/data-security/gdpr-article-32/ ↩
91. Privacy-Preserving Federated Learning – Future Collaboration and ..., accessed May 22, 2025, https://www.nist.gov/blogs/cybersecurity-insights/privacy-preserving-federated-learning-future-collaboration-and ↩
92. The privacy-explainability trade-off: unraveling the impacts of differential privacy and federated learning on attribution methods - PubMed Central, accessed May 22, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC11253022/ ↩
93. Sindhuja Madabushi, accessed May 22, 2025, https://sindhujamadabushi.github.io/ ↩
94. Empirical Analysis of Privacy-Fairness-Accuracy Trade-offs in Federated Learning: A Step Towards Responsible AI - ResearchGate, accessed May 22, 2025, https://www.researchgate.net/publication/390039157_Empirical_Analysis_of_Privacy-Fairness-Accuracy_Trade-offs_in_Federated_Learning_A_Step_Towards_Responsible_AI ↩
95. NIST Proposes Updates to Privacy Framework - Connect On Tech, accessed May 22, 2025, https://connectontech.bakermckenzie.com/nist-proposes-updates-to-privacy-framework/ ↩