Homomorphic encryption offers two extraordinary advantages: first, it has the potential to be secure against intrusions even from quantum computers; Second, it allows users to use the data for computation without requiring decryption, allowing secure data offload to commercial clouds and other external locations. However – as researchers from Intel and the Barcelona Supercomputing Center (BSC) explained – homomorphic encryption “is not free from drawbacks that currently make it impractical in many scenarios”, including “size. data increases dramatically when encrypted, ”limiting its application for large neural networks.
Now that could change: BSC and Intel have, for the first time, run large homomorphically encrypted neural networks.
“Homomorphic encryption… allows inference using encrypted data, but it incurs memory and execution overheads of 100x to 10,000x,” the authors wrote in their article. “Secured in depth neural network … inference using [homomorphic encryption] is currently constrained by compute and memory resources, with frameworks requiring hundreds of gigabytes of DRAM to evaluate small models.
To do this, the researchers deployed Intel technology: in particular, Intel Optane persistent memory and Intel Xeon Scalable processors. Optane memory has been combined with DRAM to complement the higher capacity of persistent memory with the faster speeds of DRAM. They tested the combination using a variety of different setups to run large neural networks, including ResNet-50 (now the largest neural network ever to run with homomorphic encryption) and the largest variant of MobileNetV2. As a result of the experiments, they landed on a setup with only a third of the DRAM, but only a 10% drop in performance compared to a fully DRAM system.
“This new technology will enable the widespread use of neural networks in cloud environments, including, for the first time, where unquestionable confidentiality is required for the data or the neural network model,” said Antonio J. Peña , the BSC researcher who led the study and leader of the BSC Accelerators and Communications for High Performance Computing team.
“Computing is both computationally and memory intensive,” added Fabian Boemer, technical manager at Intel who supports this research. “To accelerate the memory access bottleneck, we are studying different memory architectures that allow better near-memory computation. This work is an important first step in solving this often overlooked challenge. Among other technologies, we are studying the use [of] Intel Optane persistent memory to keep data accessible at all times near the processor during evaluation.
To learn more about this research, read the research paper, which is available in its entirety. here.