AdaPI is a novel approach designed to facilitate adaptive private inference (PI) in edge computing, addressing the challenge of varying energy budgets across different edge devices. The method optimizes a DNN model to perform well across devices with diverse energy constraints by employing a PI-aware training strategy that optimizes weight-level and feature-level soft masks. These masks are then transformed into multiple binary masks to adjust communication and computation workloads. AdaPI achieves optimal accuracy for each energy budget, outperforming state-of-the-art (SOTA) PI methods by 7.3% on the CIFAR-100 dataset. The code for AdaPI is available on GitHub. - **Model Adaptivity**: AdaPI enables a single set of model weights to adapt to varying energy budgets, reducing the need for specialized models. - **Soft Masks and Indicator Functions**: AdaPI uses soft masks and an indicator function to optimize multiple masks associated with different computation and communication workloads. - **Sequential Multi-Mask Training**: A sequential training strategy is proposed to maximize accuracy for each binary mask, ensuring efficient deployment across diverse devices. - **Unified Metric**: A normalized ReLU count metric is introduced to evaluate the trade-off between MAC and ReLU reductions, providing a comprehensive comparison with prior works. - **CIFAR-10/CIFAR-100**: AdaPI achieves significant accuracy improvements over SOTA methods, with a 7.3% increase on CIFAR-100. - **Tiny-ImageNet**: AdaPI demonstrates favorable performance compared to SOTA methods, showing adaptivity to different energy budgets. - **Energy Consumption**: AdaPI can accommodate devices with varying energy budgets by selecting suitable masks, reducing latency and communication volume. - **Model Efficiency**: AdaPI strikes a balance between accuracy and deployment efficiency, making it suitable for edge computing environments. - **U.S. National Science Foundation**: Partial funding for this work is provided by the U.S. National Science Foundation under various grants.AdaPI is a novel approach designed to facilitate adaptive private inference (PI) in edge computing, addressing the challenge of varying energy budgets across different edge devices. The method optimizes a DNN model to perform well across devices with diverse energy constraints by employing a PI-aware training strategy that optimizes weight-level and feature-level soft masks. These masks are then transformed into multiple binary masks to adjust communication and computation workloads. AdaPI achieves optimal accuracy for each energy budget, outperforming state-of-the-art (SOTA) PI methods by 7.3% on the CIFAR-100 dataset. The code for AdaPI is available on GitHub. - **Model Adaptivity**: AdaPI enables a single set of model weights to adapt to varying energy budgets, reducing the need for specialized models. - **Soft Masks and Indicator Functions**: AdaPI uses soft masks and an indicator function to optimize multiple masks associated with different computation and communication workloads. - **Sequential Multi-Mask Training**: A sequential training strategy is proposed to maximize accuracy for each binary mask, ensuring efficient deployment across diverse devices. - **Unified Metric**: A normalized ReLU count metric is introduced to evaluate the trade-off between MAC and ReLU reductions, providing a comprehensive comparison with prior works. - **CIFAR-10/CIFAR-100**: AdaPI achieves significant accuracy improvements over SOTA methods, with a 7.3% increase on CIFAR-100. - **Tiny-ImageNet**: AdaPI demonstrates favorable performance compared to SOTA methods, showing adaptivity to different energy budgets. - **Energy Consumption**: AdaPI can accommodate devices with varying energy budgets by selecting suitable masks, reducing latency and communication volume. - **Model Efficiency**: AdaPI strikes a balance between accuracy and deployment efficiency, making it suitable for edge computing environments. - **U.S. National Science Foundation**: Partial funding for this work is provided by the U.S. National Science Foundation under various grants.