Adaptive Gradient Methods for Differentially Private TinyML in 6G

The sixth-generation (6G) of wireless systems is poised to enable a hyper-connected world of intelligent devices, where tiny machine learning (TinyML) will drive pervasive, real-time applications. However, this paradigm, built on distributed data from billions of endpoints, introduces an unprecedented privacy attack surface. A fundamental challenge for deploying AI in 6G is ensuring robust data privacy on resource-constrained devices without sacrificing model utility. Differentially Private Stochastic Gradient Descent (DP-SGD), a cornerstone of private machine learning, critically depends on managing gradient sensitivity, a task traditionally hampered by the manual tuning of a static clipping threshold. This paper presents a comprehensive analysis of gradient control mechanisms for DP-SGD, evaluated from a 6G deployment perspective. We trace the evolution from static clipping to fully adaptive scaling methods that obviate the need for a fixed threshold. To unify these approaches, we propose a conceptual framework, culminating in a case study of the Differentially Private Per-sample Adaptive Scaling Clipping (DP-PSASC) algorithm. We argue that such adaptive methods are not merely algorithmic improvements but are essential, “6G-adaptive” solutions that can be integrated into next-generation network architectures, such as the Open RAN (O-RAN) framework, to deliver efficient, scalable, and trustworthy AI.