BSC Runs Large Encrypted Neural Networks for the First Time Using Intel Optane Persistent Memory and Intel Xeon Scalable Processors

In collaboration with Intel, Barcelona Supercomputing Center-Centro Nacional de Supercomputación has managed to encrypt the execution of large neural networks efficiently. They have achieved this thanks to Intel Optane persistent memory and Intel Xeon scalable processors with built-in artificial intelligence acceleration.

Until now, the main memory size supported by current technology has limited the use of homomorphic encryption to small models of neural networks designed for portable devices. This is why the encryption of large neural networks is a great technological advance. This type of encryption cannot be decrypted even by quantum computers, and allows operations to be carried out directly on the encrypted data, so that the entity that operates it does not have access to its content.

A new layer of security for unsafe environments

Since this encryption does not need to be decrypted to work, privacy in insecure environments is guaranteed, such as in the cloud. The main challenge with homomorphic encryption is its additional cost, since it increases as the size of the data increases, which can be multiplied by up to 10.000. Intel Optane persistent memory offers much greater capacities than DRAM in faster access time than other non-volatile memory.

Despite not being as fast as main memory technology, combining the two with an efficient access pattern offers huge advantages in terms of price and performance. This new technology can be applied in the private execution of neural networks in unreliable remote environments, such as the cloud. It includes both the protection of intellectual property related to the neural network model, as well as the data
used, thereby allowing compliance with numerous data protection laws and regulations in different countries. The research has been carried out by a team of BSC researchers, together with an international team from Intel, with members from both Europe and the United States.

"This new technology will enable the widespread use of neural networks in cloud environments, even, and for the first time, when unquestionable confidentiality of the data or the neural network model is required."

Antonio J. Peña, senior researcher at the BSC.

The scientific article related to this research will be published in the journal IEEE Transactions on Computers, and it analyzes the execution of the popular ResNet-50 model, which incorporates 25 million parameters and consumes almost 1TB of memory, which is more than double that of a computation node of the MareNostrum4 supercomputer. This article will also mention the architecture of an efficient computer for this task, with only a third of normal RAM, which usually consumes about 10 times more power per byte than Intel Optane persistent memory, which allows configurations with greater efficiency. energy and sustainability of the solution.