How to get the best results？

[zixuWang2022]

Hi~!
I am recently reading your paper "Generating Effective Software Obfuscation Sequences With Reinforcement Learning". After my experiments, I have successfully run the code in the paper and obtained some results. However, I would like to ask you how to get the optimal results from these results?

I would like to confirm the method of obtaining the optimal results to ensure that my experimental results are not interpreted incorrectly or arbitrarily. I would appreciate any guidance or reference you can provide.

Thank you for your time and help.
Best regards.

[whj0401]

It will output a sequence of obfuscation choices for each iteration. Given a binary and after many iterations, the output obfuscation sequence will become stable. You can store the model and remove the random mechanism to get the final optimal sequence. Or you can simply select the sample with the highest score amoung all produced binaries.

[zixuWang2022]

Hi~！
I'm coming again. First of all, thank you for your guidance in the previous question. I have one more question about the code that I need to ask you. I was reading the code in the file RLOBF/src/coreutils_gym_env.py, line 18 of the c2 variable

c2 = 0.37486390900521455
#c2 = 0.0 # here we simply ignore the overhead

I can understand by the comments that the performance overhead caused by obfuscation is ignored when c2 equals 0. But I don't understand how to limit the performance overhead caused by obfuscation when c2 is not equal to 0. For example, if I want the performance bloat to be no more than twice that of the source program, how should I set the value of c2.

[whj0401]

Please see here.
c2 is a hyperparameter that is tuned to balance the weight of time overhead.
It influences the reward function, but it cannot guarantee the impact of obfuscation on efficiency.

By tuning the value, you can make the model pay more or less attention to efficiency or similarity.
Since there are random mechanisms during obfuscation, you likely find some cases with sufficiently low
overhead. Note that the time cost is not a ``deterministic value'' which can be influenced by the execution
environment such as CPU, cache, disk, etc. It is hard to provide a guarantee. In our paper, we ran the samples
with the same inputs multiple times and computed the average overhead. It is an approximation but not a guarantee.,

[zixuWang2022]

Thank you for your reply. I have a question for you, can I modify the code here to achieve the time overhead limit?