https://ryanorban.com/categories/neurips-2021/Recent content in Neurips-2021 on Ryan OrbanHugoen-usme@ryanorban.com (Ryan Orban)me@ryanorban.com (Ryan Orban)Ryan OrbanTue, 19 Apr 2022 00:00:00 +0000https://ryanorban.com/notes/crypten-secure-mpc-meets-machine-learning/Tue, 19 Apr 2022 00:00:00 +0000me@ryanorban.com (Ryan Orban)https://ryanorban.com/notes/crypten-secure-mpc-meets-machine-learning/<h3 id="summary" class="scroll-mt-8 group"> Summary <a href="#summary" class="no-underline hidden opacity-50 hover:opacity-100 !text-inherit group-hover:inline-block" aria-hidden="true" title="Link to this heading" tabindex="-1"> <svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" class="lucide lucide-link w-4 h-4 block" viewBox="0 0 24 24" > <path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71" /> <path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71" /> </svg> </a> </h3> <p>Secure multi-party computation (MPC) lets parties compute jointly on their combined data without revealing what each party holds. The obstacle to adoption in machine learning has always been that MPC frameworks speak the language of circuit compilers and cryptographic protocols — not the language of tensors, backpropagation, and neural network modules. CrypTen from Facebook AI Research (NeurIPS 2021) fixes this by implementing an MPC computation layer that mirrors the PyTorch API almost exactly: same tensor ops, same autograd, same <code>nn.Module</code> structure. In principle you change one import and your code runs over encrypted data.</p>