Efficient Per-Example Gradient Computations in Convolutional Neural Networks

  • Gaspar Rochette ,
  • Andre Manoel ,
  • Eric W. Tramel

Theory and Practice of Differential Privacy (TPDP) workshop at CCS 2020 |

Paper was accepted to workshop after being peer-reviewed, but there were no formal proceedings.

Publication | Publication | Publication

Deep learning frameworks leverage GPUs to perform massively-parallel computations over batches of many training examples efficiently. However, for certain tasks, one may be interested in performing per-example computations, for instance using per-example gradients to evaluate a quantity of interest unique to each example. One notable application comes from the field of differential privacy, where per-example gradients must be norm-bounded in order to limit the impact of each example on the aggregated batch gradient. In this work, we discuss how per-example gradients can be efficiently computed in convolutional neural networks (CNNs). We compare existing strategies by performing a few steps of differentially-private training on CNNs of varying sizes. We also introduce a new strategy for per-example gradient calculation, which is shown to be advantageous depending on the model architecture and how the model is trained. This is a first step in making differentially-private training of CNNs practical.