AutoPrivacy: Automated Layer-wise Parameter Selection for Secure Neural Network Inference
Authors: Qian Lou, Song Bian, Lei Jiang
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We performed extensive experiments to show the consistent effectiveness of Auto Privacy to minimize the HPPNN inference latency with trivial loss of accuracy. |
| Researcher Affiliation | Academia | Qian Lou Indiana University Bloomington louqian@iu.edu Song Bian Kyoto University sbian@easter.kuee.kyoto-u.ac.jp Lei Jiang Indiana University Bloomington jiang60@iu.edu |
| Pseudocode | No | The paper describes the approach, including the DDPG agent, in textual form but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide a specific link or explicit statement indicating that the source code for the described methodology (Auto Privacy) is publicly available. |
| Open Datasets | Yes | Our experiments are performed on the CIFAR-10/100 dataset. |
| Dataset Splits | No | The paper states 'Only 7CNET is trained and tested on CIFAR-10, while experiments of RESNET and MOBNET are performed on CIFAR-100', but it does not provide specific percentages, sample counts, or detailed methodology for dataset splits (training, validation, test). |
| Hardware Specification | Yes | We ran HPPNN inferences and measured the latency of each type of operations on an Intel Xeon E7-4850 CPU with 1TB DRAM. We implemented and trained Auto Privacy on a Nvidia GTX1080-Ti GPU. |
| Software Dependencies | No | The paper mentions using 'Microsoft SEAL library [6]' and 'swanky library [23]' but does not specify their version numbers. |
| Experiment Setup | Yes | The DDPG agent is trained with fixed learning rates, i.e., 10 4 for the actor network and 10 3 for the critic network. The replay buffer size of Auto Privacy is 2000. During exploration, the DDPG agent adds a random noise to each action. The standard deviation of Gaussian action noise is initially set to 0.5. After each episode, the noise is decayed exponentially with a decay rate of 0.99. |