Curvature Clues: Decoding Deep Learning Privacy with Input Loss Curvature

Authors: Deepak Ravikumar, Efstathia Soufleri, Kaushik Roy

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our theoretical findings through experiments in computer vision classification tasks, demonstrating that input loss curvature surpasses existing methods in membership inference effectiveness.
Researcher Affiliation Academia Deepak Ravikumar Efstathia Soufleri Kaushik Roy Department of Electrical and Computer Engineering Purdue University West Lafayette, IN 47907 {dravikum, esoufler, kaushik}@purdue.edu
Pseudocode Yes The pseudo-code for obtaining input loss curvature score using zero order estimation shown in Algorithm 1 in Appendix A.1.
Open Source Code Yes Code available at https://github.com/Deepak Tatachar/Curvature-Clues
Open Datasets Yes Specifically, we use the CIFAR10, CIFAR100 [Krizhevsky et al., 2009] and Image Net [Russakovsky et al., 2015] datasets.
Dataset Splits Yes Shadow models for CIFAR10 and CIFAR100 were trained on a 50% subset of the data for 300 epochs. For Image Net, we used pre-trained models from Feldman and Zhang [2020], trained on a 70% subset of Image Net.
Hardware Specification Yes All of the experiments were performed on a heterogeneous compute cluster consisting of 9 1080Ti s, 6 2080Ti s and 4 A40 NVIDIA GPUs, with a total of 100 CPU cores and a combined 1.2 TB of main system memory.
Software Dependencies No The paper mentions using the "Opacus library [Yousefpour et al., 2021]" and cites a paper on "Py Torch 2 [Ansel et al., 2024]", but it does not provide specific version numbers for these software components or any other key libraries or dependencies used in the experiments.
Experiment Setup Yes Details regrading hyperparameters are provided in Appendix A.13. The initial learning rate was set to 0.001. The learning rate is decreased by 10 at epochs 12 and 16 with a batch size of 128. We used SGD optimizer with the initial learning rate set to 0.1, weight decay of 1 10 4 and momenutm of 0.9. The learning rate was decayed by 0.1 at 180th and 240th epoch. Our code uses 2 hyper parameters for zero-order input loss curvature estimation. The niter and h in Algorithm 1. We used niter = 10 and h = 0.001.