FULL PDF
Abstract
The proliferation of intelligent applications on microcontrollers and low power devices has underscored the urgency for privacy preserving machine learning paradigms. Following this cloud-based infrastructure paradigm, where latency, privacy, and compliance concerns arise, TinyML Machine Learning on ultra-resource constrained devices has come forth as a key solution towards decentralized intelligence. However, introducing smart computation at the edge level raises very serious privacy and regulatory concerns in sensitive fields, e.g., in healthcare, smart homes, and industrial IoT. We present here a policy aligned architectural framework for privacy-aware TinyML systems. With our approach, mechanisms ensuring policy compliance and confidentiality are imposed directly into the training and inference workflows of TinyML devices, such as through programmable consent layers, adaptive anonymization modules, and real-time compliance engines mandated by regulation. The framework is assessed across indicative scenarios, thereby showing that strong privacy guarantees can be attained without any tradeoff in computation efficiency and learning mesh. This work merges embedded intelligence with contemporary privacy governance and supplies a scalable, lawful, and ethically aligned model for TinyML system deployments within real-world settings.
View more >>