Privacy-Preserving AI on Edge Devices: A Comparative Study of On-Device Learning and Cloud Processing

The rise of edge computing has transformed how artificial intelligence (AI) processes personal and sensitive data. Traditional cloud-based AI architectures raise privacy concerns due to the transmission and centralized storage of user data. On-device learning, by contrast, enables local inference, enhancing data privacy and reducing network dependency. This study conducts a comparative analysis of on-device learning and cloud-based AI processing using a human activity recognition (HAR) task. We evaluate both approaches on model accuracy, latency, energy consumption, and privacy exposure. Our findings reveal that while cloud-based models slightly outperform in accuracy and energy efficiency, on-device models offer significantly lower latency and enhanced privacy by eliminating the need for data transmission. The results support a privacy-preserving edge AI paradigm, particularly for applications in healthcare, personal security, and mobile computing. Recommendations are provided to guide system architects in selecting deployment strategies aligned with privacy, performance, and regulatory requirements.