Privacy-Preserving Machine Learning at the Edge: A Comparative Study of Federated Learning, Differential Privacy, and Secure Aggregation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The increasing use of artificial intelligence on smartphones, Internet of Things devices, edge servers, and other distributed platforms has created new opportunities for intelligent applications but has also intensified concerns regarding the privacy of machine-learning data. Conventional centralized machine learning requires sensitive data to be collected and transferred to a central server, creating potential risks of unauthorized access, data leakage, and misuse. Privacy-Preserving Machine Learning (PPML) addresses these concerns by introducing techniques that allow useful models to be trained while reducing exposure of sensitive information. Among the most important approaches for edge environments are Federated Learning (FL), Differential Privacy (DP), and Secure Aggregation (SA). Federated Learning keeps raw training data on participating devices and transfers model updates rather than the original data [1]. Differential Privacy provides a mathematical framework for limiting the privacy loss associated with individual data records or users [2,3]. Secure Aggregation uses cryptographic techniques so that a coordinating server can obtain an aggregate of client updates without directly observing individual contributions [4]. This paper presents a comparative analysis of these three approaches with respect to privacy protection, accuracy, communication overhead, computational requirements, scalability, and suitability for edge computing. Published results indicate that federated learning can reduce communication rounds substantially compared with conventional synchronized training, while practical secure aggregation has demonstrated measurable but manageable communication expansion under representative settings [1,4]. The study further develops a multi-criteria evaluation framework and discusses the advantages of combining FL, DP, and SA rather than relying on a single privacy mechanism. The paper concludes that layered privacy protection offers a stronger foundation for trustworthy edge intelligence, although challenges involving non-IID data, client heterogeneity, privacy-utility trade-offs, malicious participants, communication costs, and deployment complexity remain significant research opportunities.