On Leveraging GPUs for Security: discussing k-anonymity and pattern matching

Recent GPU improvements led to an increased use of these hardware devices for tasks not related to graphical issues. With the advent of GPGPU, programmers have token advantage of GPUs increasingly, trying to improve a wide variety of algorithms via parallelization. According to the need to make GPUs easily programmable, the manufacturers of the graphic cards developed some parallel computing libraries such as CUDA, by NVIDIA, and OpenCL, by Khronos Group. Privacy and security problems can benefit from parallel computing resources.

Contributions
In our work we contributed to improve GPU usefulness for security, showing that it is possible to improve performance thanks to this hardware component.
Furthermore, Aparapi allowed us to extend Java’s paradigm “Write Once Run Anywhere” to parallel computing.
To reach this goal we have studied the k-anonymity problem, emphasizing its mathematical aspects but also the need for fast results calculation, in order to keep the privacy of the respondents of a large amount of datasets all over the world.
We have performed an in-depth thorough study of the actual parallelization issues of algorithms on GPUs. As a consequence, we have learnt a big deal of information over the advantages, benefits and caveats of such approach.
Particularly, we have focused on the micro-aggregation technique, investigating two algorithms, MDAV and HyperCubes. We have improved their performance and we have also proposed an hybrid version for the MDAV algorithm.
We seen that parallelization is very helpful in improving performance of algorithms which contain parallelizable portions of code. Furthermore, GPUs allow the parallelization to be very performant, even if there are still several issues related to the usage of graphic cards.
In conclusion we have collected a large amount of high-quality experimental data that can be used further to better model actual GPU usage. These GPU collected results are particularly interesting, as the trend for future GPUs and CPUs will see:
1. a widening of the gap for performance and
2. a unified address space, resulting in a scaling issues improvement.

Lessons learnt
Our work allowed us to make some observations about Aparapi and parallelization.
Despite several issues, such as the lack of support for multiple entrypoints kernel, Aparapi has shown to be a very powerful tool, that allows to fully exploit the power of the GPUs.
We have learnt how to improve performance avoiding the main bottleneck due to GPU usage: the data transfer between CPU and GPU. We have also learnt how to achieve the best performance from different hardware architectures, managing the chunkSize value.
In conclusion it is not trivial to parallelize on GPUs: the development of parallel code implementing tools is not finished yet.

Future works
After having devised a hybrid version of the MDAV algorithm, we are going to improve our approach allowing the micro-aggregation of larger datasets. In practice we are going to overcome the GPU memory limits by partitioning the dataset to be micro-aggregated and by computing micro-aggregation of different subsets on different GPUs. In this approach it is important to pay attention to how data are distributed on the GPUs, in order to avoid the increase of the SSE. A solution to this issue consists in totally ordering the records in the dataset, by one or two components, before partitioning it.