On September 2, NTT announced that it has developed a new technology that can learn deep learning while encrypting data. A technique for performing computation while encrypted is called “secret computation”. This time, the calculation processing that was not good at the conventional secret calculation can be calculated at high speed and with high accuracy.

Usually, it is necessary to decrypt the original data to calculate the encrypted data. However, since decryption increases the risk of information leakage, there is a problem that it is difficult to collect and use data due to privacy problems. With new technology, data can be calculated with encryption, so deep learning can be learned and prediction processing can be performed while privacy is protected. Since the resistance of data providers can be reduced, more training data may be collected than before.

Deep learning uses a “soft max function” that converts input to an output of 0 to 1, and “Adam” that is an optimization processing method. These perform a combination of division, exponent, reciprocal, square root, etc., but traditional secret calculations were not good at such processing. With the new technology, the softmax function can be calculated quickly and accurately, and Adam can also be used.

Deep learning has a plurality of “intermediate layers” for extracting features of the original data between the “input layer” indicating the original data and the “output layer” indicating the response. The company prepared a correspondence table with input / output pairs in order to calculate Softmax functions and Adam. The company has developed a unique technology called a secret map that encrypts the input and the correspondence table while providing an output corresponding to the input.It also developed a high-speed algorithm dedicated to each of the division, exponent, reciprocal, and square root that make up the softmax function and Adam.

According to the company, using a dedicated high-speed algorithm, it was possible to execute one epoch (learning once for training data) in just over 5 minutes by learning a model that discriminates 60,000 handwritten characters.