Detecting and Mitigating Errors in Neural Networks

In our modern world, neural networks are ubiquitous. Their functionality and complexity keep increasing while the feature size of the hardware that runs today’s neural networks keeps continuously shrinking. In addition, the number of attacks keeps continuously growing. Neural networks are interesting targets, especially when used in safety-relevant contexts such as medical devices. In this paper, we will present a framework that addresses these two issues. The framework can detect intentional and unintentional changes in the network’s architecture and its parameters. A two-step approach is used to restore the network to an unchanged state. In the first step, changes to the network are detected. The step is performed online concurrent with the normal use of the neural network. In the second step, the parameters of the network will be restored. To decrease the time required for parameter restoration advanced power analysis is used. The framework can be used for a great variety of network architectures. Special attention is taken for very fast error recovery. In addition, resource requirements, including processor power and storage space are reduced as far as possible.