Privacy-Preserving Artificial Intelligence: Principles, Methods, Applications, and Challenges

Connecting the dots in trustworthy Artificial Intelligence: From AI principles, ethics, and key requirements to responsible AI systems and regulation

It is our great pleasure to welcome you to the 10th annual Workshop on Encrypted Computing and Applied Homomorphic Cryptography - WAHC'22. WAHC was created in 2013 as a forum to organize and foster discussion of a wide variety of aspects of encrypted computing and secure computation. In a world where distance is no longer an obstacle for cooperation, secure computation is becoming a key feature of current and future information systems. Distributed network applications and cloud architectures are at danger because lots of personal consumer data is aggregated in all kinds of formats and for various purposes. Industry and consumer electronics companies are facing massive threats like theft of intellectual property and industrial espionage. Public infrastructure has to be secured against sabotage and manipulation. A possible solution is encrypted computing: Data can be processed on remote, possibly insecure resources, while program code and data is encrypted all the time. This allows to outsource the computation of confidential information independently from the trustworthiness or the security level of the remote system. The goal of the workshop is to bring together researchers with practitioners and industry to present, discuss and to share the latest progress in the field. We want to exchange ideas that address real-world problems with practical approaches and solutions. The workshop is typically uniformly attended by academia, government, and industry, with attendees both from prior years with experience in the domain and new attendees learning from the community. Specific encrypted computing technologies focus on homomorphic encryption and secure multiparty computation but also need to interoperate with existing technologies and infrastructures. The technologies and techniques discussed in this workshop are key to extending the range of applications that can be securely and practically outsourced.

The exponential growth of data in recent years has created unprecedented opportunities for insights through big data analytics. However, this surge in data generation also raises significant privacy concerns, especially when sensitive information is involved. Privacy-preserving techniques have emerged as a critical area of research to ensure the ethical use of data while maintaining analytical efficacy. This paper explores state-of-the-art approaches to safeguarding privacy in big data analytics. Techniques such as data anonymization, differential privacy, secure multi-party computation, and homomorphic encryption are examined for their effectiveness and applicability across various domains. Anonymization techniques aim to remove personally identifiable information while retaining analytical utility, though challenges like re-identification attacks persist. Differential privacy introduces calibrated noise to data, balancing privacy and accuracy. Secure multi-party computation enables collaborative analytics without sharing raw data, promoting confidentiality in distributed environments. Homomorphic encryption allows computations on encrypted data, ensuring security throughout the analytical process. The integration of privacy-preserving techniques into big data workflows demands addressing computational overheads, scalability, and regulatory compliance. This paper also highlights how combining multiple privacy-preserving methods can mitigate individual limitations and enhance overall robustness. As privacy regulations evolve, such as the General Data Protection Regulation (GDPR) and similar frameworks, these techniques become increasingly vital.

The rapid evolution of artificial intelligence (AI) has transformed how personal data is harnessed, driving innovation in diverse fields, from healthcare to finance. However, this transformation brings complex challenges in safeguarding individual privacy while leveraging vast amounts of personal information. This paper examines the dual role of AI as both a catalyst for technological advancement and a potential risk to privacy. By analyzing current trends in AI data usage, exploring the inherent challenges in balancing privacy and functionality, and highlighting emerging solutions—such as federated learning and differential privacy—the study offers a comprehensive overview of strategies for responsible AI development. The findings underscore the need for ethical frameworks, robust regulatory measures, and innovative privacy-preserving techniques to ensure that the benefits of AI do not come at the expense of personal data protection. Index Terms— Artificial Intelligence (AI), Personal Data, Privacy, Data Protection, Privacy-Preserving Machine Learning, Federated Learning, Differential Privacy, Ethical AI, Data Governance, Regulatory Compliance.

IntroductionThe integration of artificial intelligence (AI) into medical text generation is transforming public health by enhancing clinical documentation, patient education, and decision support. However, the widespread deployment of AI in this domain introduces significant ethical challenges, including fairness, privacy protection, and accountability. Traditional AI-driven medical text generation models often inherit biases from training data, resulting in disparities in healthcare communication across different demographic groups. Moreover, ensuring patient data confidentiality while maintaining transparency in AI-generated content remains a critical concern. Existing approaches either lack robust bias mitigation mechanisms or fail to provide interpretable and privacy-preserving outputs, compromising ethical compliance and regulatory adherence.MethodsTo address these challenges, this paper proposes an innovative framework that combines privacy-preserving AI techniques with interpretable model architectures to achieve ethical compliance in medical text generation. The method employs a hybrid approach that integrates knowledge-based reasoning with deep learning, ensuring both accuracy and transparency. Privacy-enhancing technologies, such as homomorphic encryption and secure multi-party computation, are incorporated to safeguard sensitive medical data throughout the text generation process. Fairness-aware training protocols are introduced to mitigate biases in generated content and enhance trustworthiness.Results and discussionThe proposed approach effectively addresses critical challenges of bias, privacy, and interpretability in medical text generation. By combining symbolic reasoning with data-driven learning and embedding ethical principles at the system design level, the framework ensures regulatory alignment and improves public trust. This methodology lays the groundwork for broader deployment of ethically sound AI systems in healthcare communication.

The integration of Internet of Things (IoT) technologies within smart healthcare systems has revolutionized medical diagnostics, patient monitoring, and treatment personalization, the widespread deployment of interconnected medical devices introduces a complex landscape of security and privacy risks, exposing sensitive health data to cyber threats and unauthorized access. As healthcare systems generate vast volumes of real-time data across distributed environments, traditional security mechanisms fall short in ensuring end-to-end protection, especially under the resource constraints of IoT devices. Artificial intelligence (AI) has emerged as a transformative force, enabling proactive threat detection, intelligent risk scoring, and adaptive mitigation strategies that operate in dynamic healthcare ecosystems. This chapter presents a comprehensive analysis of IoT security architectures in smart healthcare, examining vulnerabilities at each system layer and highlighting AI-driven frameworks for intrusion detection, anomaly recognition, and privacy-preserving data analytics. Advanced methods such as federated learning, differential privacy, homomorphic encryption, and secure multi-party computation are explored as essential tools for safeguarding patient information while maintaining model utility and interoperability. The convergence of AI and cybersecurity offers a resilient approach to managing evolving threats without compromising clinical efficiency or regulatory compliance. Key challenges, including scalability, explainability, and cross-border data governance, are critically assessed, along with future directions for building robust, ethical, and intelligent healthcare infrastructures.

The emerging capabilities of artificial intelligence (AI) and the systems that employ them have reached a point where they are integrated into critical decision-making processes, making it paramount to change and adjust how they are evaluated, monitored, and governed. For this reason, trustworthy AI (TAI) has received increased attention lately, primarily aiming to build trust between humans and AI. Due to the far-reaching socio-technical consequences of AI, organisations and government bodies have already started implementing frameworks and legislation for enforcing TAI, such as the European Union’s AI Act. Multiple approaches have evolved around TAI, covering different aspects of trustworthiness that include fairness, bias, explainability, robustness, accuracy, and more. Moreover, depending on the AI models and the stage of the AI system lifecycle, several methods and techniques can be used for each trustworthiness characteristic to assess potential risks and mitigate them. Deriving from all the above is the need for comprehensive tools and solutions that can help AI stakeholders follow TAI guidelines and adopt methods that practically increase trustworthiness. In this paper, we formulate and propose the Trustworthiness Optimisation Process (TOP), which operationalises TAI and brings together its procedural and technical approaches throughout the AI system lifecycle. It incorporates state-of-the-art enablers of trustworthiness such as documentation cards, risk management, and toolkits to find trustworthiness methods that increase the trustworthiness of a given AI system. To showcase the application of the proposed methodology, a case study is conducted, demonstrating how the fairness of an AI system can be increased.

Artificial Intelligence (AI) is rapidly changing every aspect of our society—including amplifying our biases. Fairness, trust and ethics are at the core of many of the issues underlying the implications of AI. Despite this, research on AI with relation to fairness, trust and ethics in the information systems (IS) field is still scarce. This panel brought together academia, business and government perspectives to discuss the challenges and identify potential solutions to address such challenges. This panel report presents eight themes based around the discussion of two questions: (1) What are the biggest challenges to designing, implementing and deploying fair, ethical and trustworthy AI?; and (2) What are the biggest challenges to policy and governance for fair, ethical and trustworthy AI? The eight themes are: (1) identifying AI biases; (2) drawing attention to AI biases; (3) addressing AI biases; (4) designing transparent and explainable AI; (5) AI fairness, trust, ethics: old wine in a new bottle?; (6) AI accountability; (7) AI laws, policies, regulations and standards; and (8) frameworks for fair, ethical and trustworthy AI. Based on the results of the panel discussion, we present research questions for each theme to guide future research in the area of human–computer interaction.

Federated Learning (FL) is a decentralised approach to machine learning that enables model training on local devices without the need to share raw data. While FL inherently provides some level of privacy, significant challenges remain in fully protecting sensitive information. This article explores advanced artificial intelligence (AI) techniques for improving privacy in federated learning. I propose novel methods that include differential privacy, homomorphic encryption, and secure multiparty computation, enhanced by AI-driven optimizations. My empirical studies and theoretical analyses demonstrate the effectiveness and efficiency of these techniques in maintaining data protection.

Personal data is highly vulnerable to security exploits, spurring moves to lock it down through encryption, to cryptographically ‘cloud’ it. But personal data is also highly valuable to corporations and states, triggering moves to unlock its insights by relocating it in the cloud. We characterise this twinned condition as ‘clouded data’. Clouded data constructs a political and technological notion of privacy that operates through the intersection of corporate power, computational resources and the ability to obfuscate, gain insights from and valorise a dependency between public and private. First, we survey prominent clouded data approaches (blockchain, multiparty computation, differential privacy, and homomorphic encryption), suggesting their particular affordances produce distinctive versions of privacy. Next, we perform two notional code-based experiments using synthetic datasets. In the field of health, we submit a patient’s blood pressure to a notional cloud-based diagnostics service; in education, we construct a student survey that enables aggregate reporting without individual identification. We argue that these technical affordances legitimate new political claims to capture and commodify personal data. The final section broadens the discussion to consider the political force of clouded data and its reconstitution of traditional notions such as the public and the private.

This paper addresses the growing necessity of protecting AI models and their underlying data using cryptographic techniques. As AI continues to integrate into critical industries such as healthcare, finance, and autonomous systems, unique vulnerabilities arise that expose models to threats like data poisoning, adversarial attacks, and model inversion. By applying cryptographic methods like homomorphic encryption, differential privacy, and secure multiparty computation, organizations can safeguard both the integrity and confidentiality of AI models and training data. This white paper provides insights into these cryptographic approaches, detailing how each can protect against specific threats while maintaining model performance and compliance.

In recent years, to surmount the challenges of data sensitivity and data silos, federated learning (FL), a collaborative decentralized privacy preservation machine learning framework was proposed and rapidly became a current research hotspot. In this paper, we present an overview of FL workflow, as well as the communication overhead and privacy leakage problems associated with FL. In order to alleviate these two pain points, several corresponding solutions are presented in this paper along with their advantages and disadvantages. To reduce the communication overhead, this paper suggests three reliable approaches by reviewing related research including increasing the amount of computing on the edge side, model compression, and improving parallelism. To address the privacy concerns, this paper introduces three prevailing privacy preservation techniques: differential privacy, homomorphic encryption, and secure multiparty computation (SMC), and shows how to efficiently incorporate these privacy-preserving techniques into a federated learning framework by summarizing relevant literature. On the whole, this paper contributes to explaining the overall workflow of a FL framework and helps advance the understanding of FL in improving communication efficiency and enhancing privacy security.

Private approximate nearest neighbor search for on-chain data based on locality-sensitive hashing

Risk analysis is most effective when banks, insurers, and asset managers combine insights, but they're reluctant to share raw data due to confidentiality concerns. We propose a federated graph learning framework using ST-GNNs with differential privacy, homomorphic encryption, and secure multiparty computation. Institutions co-create shared risk views without revealing client or transaction information. Experiments demonstrate significant outperformance over centralized and federated baselines, achieving 0.891 AUC under strongest privacy constraints. Critically, it provides 78% early warning—a full month before traditional triggers—with only 4.7% accuracy degradation under aggressive data-shielding. This enables actionable, confidential oversight of systemic risks for regulators and enterprises.RetryClaude can make mistakes. Please double-check responses.

Efficient privacy-preserving aggregation for demand side management of residential loads