GPU implementation of RBF neural networks in audio steganalysis
Authors
Abstract
The Graphics Processing Unit (GPU) has become an integral part of today?s mainstream computing systems. First used as processors to accelerate graphics rendering on screen, the enormous computational potential of GPUs has led to the research in general-purpose computation on GPUs (GPGPU), using GPUs for non-graphics computations such as image processing, neural networks (NNs), linear algebra, sorting, computational physics and database queries (1).Over the past few years, the raw computational power of GPUs has surpassed by far that of top range CPUs (2). Unlike general-purpose processors, GPUs are optimized to perform floating-point operations, in parallel, on large data sets using
the paradigm Single Instruction Multiple Data (SIMD). In the field of machine learning there are several algorithms that can benefit from a parallel approach.
Unfortunately, most implementations are closed source or never made available to other researchers. Aproposed GPU machine learning library (GPUMLib), aims to offer an high optimized, open source, machine learning library, with several algorithms already implemented.
In this thesis we present the GPUMLib and implement a new module featuring radial basis function networks. We explore two learning models, batch and
incremental learning, both implemented in CUDA.We feature the batch learning with Radial Basis Functions Networks (RBFN), and the incremental learning with
Resource Allocating Networks with Long-Term Memory(RAN-LTM). The implementations were tested to assess its correctness and performance. Using datasets from the UCI (4) benchmarks we obtain accuracies up to 97% and speedups up to 10x.
Furthermore, weexplore the real-world application of audio steganalysis.
With steganography techniques, the art of concealed writing, we can hide information in unsuspected sources, like images, video and audio. Steganalysis aims to detect and recover hidden messages from tampered media. We applied our application to the detection of hidden messages in audio files. The results obtained showed high accuracy ratings, of up to 97%, and the CUDA implementation presented speedups of up to 15x. These experiments allowed us to draw some conclusions on the best feature extraction methods, and steganography applications.