"Improving Security as Artificial Intelligence Moves to Smartphones"

Devices such as smartphones, security cameras, and speakers, will soon rely more on artificial intelligence to increase the speed at which speech and images are processed. A compression technique, called quantization, reduces the size of deep learning models in order to lessen computation and energy costs. However, compressed AI models have been found to be more vulnerable to adversarial attacks that could cause models to misclassify altered images. MIT and IBM researchers have developed a technique to improve the security of compressed AI models against such attacks. This article continues to discuss findings of a new study conducted by MIT and IBM researchers in relation to the vulnerability of compressed deep learning models to adversarial attacks and the technique developed to reduce this vulnerability. 

MIT News report "Improving Security as Artificial Intelligence Moves to Smartphones"