Federated Learning (FL) enables collaborative model training among clients without centralising data, making it a widely adopted privacy enhancing technology (PET). Despite its privacy benefits, FL remains vulnerable to orchestrator-driven privacy attacks. In this paper, we study an underexplored threat in which a dishonest orchestrator intentionally manipulates the aggregation process to induce targeted overfitting in local models of specific clients. Although prior work focuses on reducing information leakage during training, we emphasise early client-side detection of targeted overfitting, allowing clients to disengage before significant harm occurs. To this end, we propose three detection techniques—label flipping, backdoor trigger injection, and model fingerprinting—which enable clients to verify the integrity of the global aggregation. We evaluated our methods across multiple datasets and attack scenarios. In single-client attacks, all three methods detect orchestrator-induced overfitting within 1–2 training rounds with F1 scores exceeding 0.7. Scalability experiments further show that detection effectiveness is influenced by cohort composition and method parameters. These results demonstrate that client-side integrity testing can provide early, effective, and scalable detection, supporting safer deployment of FL systems.

Federated Learning (FL) is a crucial approach for training large-scale AI models while preserving data locality, eliminating the need for centralised data storage. In collaborative learning settings, ensuring data quality is essential, and in FL, maintaining privacy requires limiting the knowledge accessible to the central orchestrator, which evaluates and manages client contributions. Accurately measuring and regulating the marginal impact of each client’s contribution needs specialised techniques. This work examines the relationship between one such technique—Shapley Values—and a client’s vulnerability to Membership inference attacks (MIAs). Such a correlation would suggest that the contribution index could reveal high-risk participants, potentially allowing a malicious orchestrator to identify and exploit the most vulnerable clients. Conversely, if no such relationship is found, it would indicate that contribution metrics do not inherently expose information exploitable for powerful privacy attacks. Our empirical analysis in a cross-silo FL setting demonstrates that leveraging contribution metrics in federated environments does not substantially amplify privacy risks.

The wide adoption of Machine Learning to solve a large set of real-life problems came with the need to collect and process large volumes of data, some of which are considered personal and sensitive, raising serious concerns about data protection. Privacy-enhancing technologies (PETs) are often indicated as a solution to protect personal data and to achieve a general trustworthiness as required by current EU regulations on data protection and AI. However, an off-the-shelf application of PETs is insufficient to ensure a high-quality of data protection, which one needs to understand. This work systematically discusses the risks against data protection in modern Machine Learning systems taking the original perspective of the data owners, who are those who hold the various data sets, data models, or both, throughout the machine learning life cycle and considering the different Machine Learning architectures. It argues that the origin of the threats, the risks against the data, and the level of protection offered by PETs depend on the data processing phase, the role of the parties involved, and the architecture where the machine learning systems are deployed. By offering a framework in which to discuss privacy and confidentiality risks for data owners and by identifying and assessing privacy-preserving countermeasures for machine learning, this work could facilitate the discussion about compliance with EU regulations and directives. We discuss current challenges and research questions that are still unsolved in the field. In this respect, this paper provides researchers and developers working on machine learning with a comprehensive body of knowledge to let them advance in the science of data protection in machine learning field as well as in closely related fields such as Artificial Intelligence.

Machine learning(ML) tools are among the promising data-driven techniques that can help solve many real-life problems. However these tools rely on the collection of large volumes of data, which raises many privacy concerns and more broadly trustworthiness concerns. Privacy Preserving technologies aim at solving the issue by integrating privacy enhancing technologies (PETs) within the machine learning pipelines.