Enabling Privacy-Preserving Collective Intelligence: A Twin In-Memory Encryption/Processing Macro Featuring Group Differential Privacy and Spatial-Temporal Ensemble

Privacy protection is a critical concern in AI, particularly in collective AI systems such as AIoT networks and federated learning. To address correlation attacks involving multiple data sources, this study presents a 22 nm twin in-memory encryption/processing (TIME) macro that integrates group differential privacy (GDP) and spatial/temporal ensemble (STE) techniques. This design enables efficient privacy protection while maintaining high accuracy, achieving 82% prediction accuracy for federated vital signs monitoring at a privacy protection strength of 2.3. With a 1Mb capacity, the system delivers a 10.49 TOPS throughput and achieves an energy efficiency of 5827 TOPS/W, highlighting its strong potential for privacy-preserving collective AI applications.