Task-aware Privacy Preservation for Multi-dimensional Data

Authors: Jiangnan Cheng, Ao Tang, Sandeep Chinchali

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our task-aware approach significantly improves ultimate task accuracy compared to standard benchmark LDP approaches with the same level of privacy guarantee. We validate the effectiveness of our task-aware approach through three real-world experiments.
Researcher Affiliation Academia 1School of Electrical and Computer Engineering, Cornell University, Ithaca, NY 2Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX. Correspondence to: Jiangnan Cheng <jc3377@cornell.edu>, Ao Tang <atang@cornell.edu>, Sandeep Chinchali <sandeepc@utexas.edu>.
Pseudocode Yes Algorithm 1 Task-aware Algorithm for ϵ-LDP Preservation in General Settings
Open Source Code Yes Our code is publicly available at https://github.com/chengjiangnan/task_aware_privacy.
Open Datasets Yes Three applications and corresponding datasets from the standard UCI Machine Learning Repository (Dua & Graff, 2017) are considered: mean estimation of hourly household power consumption, real estate valuation, and breast cancer detection. UCI machine learning repository, 2017. URL http://archive.ics.uci.edu/ml.
Dataset Splits No Table 1. Evaluation Details APPLICATION NUM OF SAMPLES TRAIN/TEST SPLIT TRAINING EPOCHS RUNTIME HOUSEHOLD POWER 1417 0.7/0.3 NA < 1 MIN REAL ESTATE 414 0.7/0.3 2000 < 2 HRS BREAST CANCER 569 0.7/0.3 2000 < 2 HRS. The paper specifies a 0.7/0.3 train/test split but does not mention a separate validation split.
Hardware Specification Yes Our evaluation runs on a personal laptop with 2.7 GHz Intel Core I5 processor and 8-GB 1867 MHz DDR3 memory.
Software Dependencies No Our code is based on Pytorch. We use the Adam optimizer and learning rate 10 3 for all the applications. The paper mentions software used (Pytorch, Adam optimizer) but does not provide specific version numbers for these dependencies.
Experiment Setup Yes We use the Adam optimizer and learning rate 10 3 for all the applications. The number of samples, train/test split, training epochs, and resulting runtime are summarized in Table 1. For task function f, we use a one-hidden-layer feedforward neural network with input size n, hidden size 1.5n and output size 1... The activation function used by the hidden layer and output layer is a Rectified Linear Unit (Re LU). For the encoder/decoder, we use a one-layer neural network (linear model)... We use one-hidden-layer feedforward neural network with input size n, hidden size n and output size n... The activation functions used by the hidden layer and output layer are a logistic and identity function, respectively. For the gradient-based learning algorithm... we set η = 0.2 and η = 0.001... and in both experiments, for each epoch we update θe and θd by 15 steps.