GPU memory leveraged for accelerated training using Tensorflow

Machine learning has been a detection technique used by many security vendors for some time now. With the enhancement brought by GPUs, many security products can now use different deep learning methods and forms of neural networks for malware classification. However, these new methods, as powerful as they are, are also limited by the amount of memory a GPU has or by the constant need of transferring data from CPU to GPU. As training for models used in security industry requires very large databases, consisting of millions of malicious and benign samples, security vendors had to look for ways to overcome memory constraints. This paper addresses this problem and presents some approaches that can be used when dealing with deep learning algorithms in conjunction with large databases, approaches that are adapted to different known machine learning frameworks like Theano or Tensorflow. The results obtained show that training time can be reduced by a factor of 30 if memory is used efficiently.