AdaPI is a novel approach for achieving adaptive private inference (PI) in edge computing, enabling a model to perform well across edge devices with varying energy budgets. The method introduces a PI-aware training strategy that optimizes model weights alongside weight-level and feature-level soft masks. These soft masks are transformed into multiple binary masks to adjust communication and computation workloads. Through sequential training with increasingly dense binary masks, AdaPI achieves optimal accuracy for each energy budget, outperforming state-of-the-art PI methods by 7.3% in test accuracy on CIFAR-100. The code is available at https://github.com/jiahuiiiiiii/AdaPI. The paper addresses the challenge of adapting DNN models to diverse edge devices with varying energy budgets, which is critical for efficient PI in edge computing. Existing PI methods are designed for static resource constraints, leading to inefficient deployment as they require specialized models for different devices. AdaPI enables a single model to adapt to multiple energy budgets by optimizing both feature-level and weight-level soft masks. These masks are converted into binary masks, allowing the model to be deployed on edge devices with varying energy constraints. AdaPI introduces a triple optimization problem to balance accuracy, computation workload, and communication workload. It also proposes a sequential multi-mask training strategy to maximize accuracy for each mask without requiring additional fine-tuning. The method uses a soft mask with an indicator function to address the challenge of optimizing multiple masks without interference. The indicator function converts optimized soft masks into binary masks based on desired densities associated with different energy levels. The proposed AdaPI method achieves a unified metric for evaluating model performance under triple optimization, converting MACs into ReLU counts for comparison. The method is evaluated on CIFAR-10, CIFAR-100, and TinyImageNet datasets, demonstrating superior performance in terms of test accuracy and normalized ReLU count. On CIFAR-100, AdaPI achieves a 7.3% improvement in accuracy compared to the state-of-the-art method SNL. The results show that AdaPI can efficiently accommodate edge devices with varying energy budgets, achieving high accuracy even under stringent resource constraints. The method is also effective on TinyImageNet, where WideResNet-22-8 outperforms ResNet-18 in terms of accuracy and efficiency. Overall, AdaPI provides a practical solution for efficient private inference in edge computing by enabling model adaptivity to diverse energy budgets.AdaPI is a novel approach for achieving adaptive private inference (PI) in edge computing, enabling a model to perform well across edge devices with varying energy budgets. The method introduces a PI-aware training strategy that optimizes model weights alongside weight-level and feature-level soft masks. These soft masks are transformed into multiple binary masks to adjust communication and computation workloads. Through sequential training with increasingly dense binary masks, AdaPI achieves optimal accuracy for each energy budget, outperforming state-of-the-art PI methods by 7.3% in test accuracy on CIFAR-100. The code is available at https://github.com/jiahuiiiiiii/AdaPI. The paper addresses the challenge of adapting DNN models to diverse edge devices with varying energy budgets, which is critical for efficient PI in edge computing. Existing PI methods are designed for static resource constraints, leading to inefficient deployment as they require specialized models for different devices. AdaPI enables a single model to adapt to multiple energy budgets by optimizing both feature-level and weight-level soft masks. These masks are converted into binary masks, allowing the model to be deployed on edge devices with varying energy constraints. AdaPI introduces a triple optimization problem to balance accuracy, computation workload, and communication workload. It also proposes a sequential multi-mask training strategy to maximize accuracy for each mask without requiring additional fine-tuning. The method uses a soft mask with an indicator function to address the challenge of optimizing multiple masks without interference. The indicator function converts optimized soft masks into binary masks based on desired densities associated with different energy levels. The proposed AdaPI method achieves a unified metric for evaluating model performance under triple optimization, converting MACs into ReLU counts for comparison. The method is evaluated on CIFAR-10, CIFAR-100, and TinyImageNet datasets, demonstrating superior performance in terms of test accuracy and normalized ReLU count. On CIFAR-100, AdaPI achieves a 7.3% improvement in accuracy compared to the state-of-the-art method SNL. The results show that AdaPI can efficiently accommodate edge devices with varying energy budgets, achieving high accuracy even under stringent resource constraints. The method is also effective on TinyImageNet, where WideResNet-22-8 outperforms ResNet-18 in terms of accuracy and efficiency. Overall, AdaPI provides a practical solution for efficient private inference in edge computing by enabling model adaptivity to diverse energy budgets.