4 ms·

How about this for an idea: a decompiler that uses Machine Learning to name the decompiled variables and functions. Would be nice even if it worked only sometim

by kubb 5y ago

How about this for an idea: a decompiler that uses Machine Learning to name the decompiled variables and functions. Would be nice even if it worked only sometimes.

planede 5y ago
It's also something that could be somewhat easy to get a lot of learning material for.
- efferifick 5y ago
  Yes, but the problem is learning through the graph data. I.e., it is not a well-structured problem like an image which will always have a certain size in pixels. Coming from a compiler background I am genuinely excited about the research on learning how to label nodes and edges in graphs, but I know little about the challenges in the ML side to bring this kind of technology into reality.
kuroguro 5y ago
For an AI-less solution there is IDA's Lumina which works pretty well. There's also a reverse engineered server for it [0] so you could build plugins for other disasemblers/decompilers to use with non-official servers. It basically hashes machine code (with address parts removed) [1], then when reverse engineers label and push symbols to the server (or get them from some debug build), others can pull and see what the functions are called in completely unrelated projects, that use the same libraries / have the same functions. [0] https://abda.nl/lumen/ https://abda.nl/lumen/ [1] https://github.com/naim94a/lumen/issues/2 https://github.com/naim94a/lumen/issues/2
efferifick 5y ago
Something very similar: a "decompiler" that decompiles minified javascript and guesses variables and functions: http://www.jsnice.org/ http://www.jsnice.org/
glouwbug 5y ago
It's certainly possible - Compile all the C projects on github with `gcc -O0`. Map statements, blocks, or functions to ASM output. Put everything in a giant SQL. Repeat for all of gcc's compiler flags. Wait, did I say it was possible? I'm curious what a neural netted compiler would produce. Probably your average CRUD software.
- azinman2 5y ago
  The thing is a classifier is going to classify. So it’ll create _some_ name… but will it be correct? Useful? Or just completely misleading? I think the misleading bits can be worse than no name at all. Better would be a high level description of what’s happening. I have a feeling that would be easier to achieve.
Andoryuuta 5y ago
I'm surprised nobody has mentioned DIRE[0] yet. They did exactly this and got some very impressive results. [0]: https://arxiv.org/abs/1909.09029 https://arxiv.org/abs/1909.09029 / J. Lacomis et al., "DIRE: A Neural Approach to Decompiled Identifier Naming," 2019 34th IEEE/ACM International Conference on Automated Software Engineering (ASE), 2019, pp. 628-639, doi: 10.1109/ASE.2019.00064. [1]: https://github.com/pcyin/dire https://github.com/pcyin/dire
- efferifick 5y ago
  This is awesome! Thanks for sharing!

by kubb 5y ago

How about this for an idea: a decompiler that uses Machine Learning to name the decompiled variables and functions. Would be nice even if it worked only sometimes.

planede 5y ago

It's also something that could be somewhat easy to get a lot of learning material for.

efferifick 5y ago
Yes, but the problem is learning through the graph data. I.e., it is not a well-structured problem like an image which will always have a certain size in pixels. Coming from a compiler background I am genuinely excited about the research on learning how to label nodes and edges in graphs, but I know little about the challenges in the ML side to bring this kind of technology into reality.

kuroguro 5y ago

For an AI-less solution there is IDA's Lumina which works pretty well. There's also a reverse engineered server for it [0] so you could build plugins for other disasemblers/decompilers to use with non-official servers. It basically hashes machine code (with address parts removed) [1], then when reverse engineers label and push symbols to the server (or get them from some debug build), others can pull and see what the functions are called in completely unrelated projects, that use the same libraries / have the same functions. [0] https://abda.nl/lumen/ https://abda.nl/lumen/ [1] https://github.com/naim94a/lumen/issues/2 https://github.com/naim94a/lumen/issues/2

efferifick 5y ago

Something very similar: a "decompiler" that decompiles minified javascript and guesses variables and functions: http://www.jsnice.org/ http://www.jsnice.org/

glouwbug 5y ago

It's certainly possible - Compile all the C projects on github with `gcc -O0`. Map statements, blocks, or functions to ASM output. Put everything in a giant SQL. Repeat for all of gcc's compiler flags. Wait, did I say it was possible? I'm curious what a neural netted compiler would produce. Probably your average CRUD software.

azinman2 5y ago
The thing is a classifier is going to classify. So it’ll create _some_ name… but will it be correct? Useful? Or just completely misleading? I think the misleading bits can be worse than no name at all. Better would be a high level description of what’s happening. I have a feeling that would be easier to achieve.

Andoryuuta 5y ago

I'm surprised nobody has mentioned DIRE[0] yet. They did exactly this and got some very impressive results. [0]: https://arxiv.org/abs/1909.09029 https://arxiv.org/abs/1909.09029 / J. Lacomis et al., "DIRE: A Neural Approach to Decompiled Identifier Naming," 2019 34th IEEE/ACM International Conference on Automated Software Engineering (ASE), 2019, pp. 628-639, doi: 10.1109/ASE.2019.00064. [1]: https://github.com/pcyin/dire https://github.com/pcyin/dire

efferifick 5y ago
This is awesome! Thanks for sharing!