Private Multiparty Perception for Navigation
Authors: Hui Lu, Mia Chiquier, Carl Vondrick
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On a new navigation dataset that we will publicly release, experiments show that private multiparty representations allow navigation through complex scenes and around obstacles while jointly preserving privacy. Section 4 analyzes the performance and efficiency of our approach on a new navigation dataset. 4 Experiments. Table 1: Plaintext and ciphertext test accuracy. |
| Researcher Affiliation | Academia | Hui Lu, Mia Chiquier, Carl Vondrick Department of Computer Science Columbia University New York, NY 10027 {hl3231, mac2500, cv2428}@columbia.edu |
| Pseudocode | No | No pseudocode or clearly labeled algorithm block was found in the paper. |
| Open Source Code | Yes | we will publicly release all code, data, and models. We call our method Cipher Nav. https://visualmpc.cs.columbia.edu |
| Open Datasets | Yes | On a new navigation dataset that we will publicly release, experiments show that private multiparty representations allow navigation through complex scenes and around obstacles while jointly preserving privacy. To analyze our framework, we construct a new dataset, The Obstacle World, which we generated based on the open-source Gym Mini World environment [7]. [7] Maxime Chevalier-Boisvert. Miniworld: Minimalistic 3d environment for rl and robotics research. https://github.com/maximecb/gym-miniworld, 2018. we will publicly release all code, data, and models. |
| Dataset Splits | No | We use 15, 000 environments for the training set, and 2, 250 environments for the testing set, which is disjoint. The training and validation datasets are also engineered to be balanced. While 'validation datasets' is mentioned, no specific split percentage or count for it is provided for reproducibility. |
| Hardware Specification | No | The plaintext models are trained on a single GPU for 1.5 days. The ciphertext 2-party computation models are trained on a single GPU for 3 days. Table 2: We show inference time comparison between ciphertext and plaintext models. Runtime (seconds) Experiment GPU CPU. While GPU and CPU are mentioned, no specific model numbers or types are provided for reproducibility. |
| Software Dependencies | No | We implement Cipher Nav in Py Torch [24] and use the Crypten [16] framework for privacy-preserving neural network operations. Specific version numbers for PyTorch and Crypten are not provided. |
| Experiment Setup | Yes | Each model is trained with 600 epochs and 0.01 learning rate. The batch size equals to 500 for ciphertext models and 100 for plaintext models. The network encoders include the view encoder E and map encoder Eφ. The view encoder E consists of two layers of convolutional layers followed by two linear layers, with Re LU activations in between. E takes in a 45 x 60 dimensional input view image and transform it into 32-dimensional vector. The map encoder Eφ consists of 3 linear layers with Re LU activations. The map encoder takes in a discrete square map represented by 5x5 matrix, and outputs map features as a 128-dimensional vector. The view image features and map features are passed into action classification network G, which is another multi-layer perceptron with four linear layers and Re LU activations. We use the mean-squared loss function to train the model F in ciphertext. |