This is a nice idea, and one I also advocate for, however it's important to keep in mind that the idea of reproducibility relies on determinism. So much of what goes into a build pipeline is inherently nondeterministic, because we're making decisions at compile time which can differ from compilation run to compilation run, setting aside flags. In fact, that's the point of an optimizing compiler, as many reproducible build projects have discovered, turning on optimizations pretty much guarantees no reproducibility.

As long as the compiler is not optimizing by "let's give this 3 seconds of solving time, then continue if no better solution is found", then optimizing is not inherently nondeterministic.

Counterpoint: Archlinux is 89% reproducible with optimizations enabled. The only thing I see which is difficult to make reproducible is optimizations with a timeout.

Instead of using a timeout, an optimization that can must be cut off if the cost is excessive can keep some kind of operation or size count, where the count is strictly a function of the input. For example, an optimization based on binary decision diagrams (BDDs) can put a ceiling on the number of nodes in the BDD.

This is defeatist: compilers do not usually use the system RNG to make decisions, so what's happening is entirely accidental introduction of difference which propagates.

There is "intentional input" (contents of the source files), and "accidental input" (source file full paths, timestamps, layout of memory given to you by the OS, and so on). A reproducible build system should give the same output for the same "intentional input".

(the only place where you do see RNG driven optimization is things like FPGA routing, which is a mess of closed toolchains anyway. It has no place in regular software compilers.)

Why does an optimizing compiler introduce nondeterminism?

In my mind an optimizing compiler is a pure function that takes source code and produces an object file.

Well, lot of things can influence here. Multithreaded build, PGO, or even the different access order of the hash table inside the code optimizer can be a factor. Things are getting probalistic and thus somewhat nondeterministic: the build itself is nondeterministic but the runtime/final execution is deterministic

"Reproducible" isn't necessary for "not modified from what everyone else gets", and that still makes some attacks FAR harder (and easier to identify, as you know what the "normal" one is). And a published Merkle tree just makes it easier to verify "none of this has changed", as opposed to SHAs on a website that could change any time.

For sure, which is one of the big benefits of git + git tagging, but the issue is even if you know you received the same binary as someone else, without reproducible and auditable builds, you have no idea if that binary originated from the same code in the case of a targeted attack.

> For sure, which is one of the big benefits of git + git tagging

That's not enough for serious security though, because git is (still) using SHA1 instead of SHA256. You would need something extra, like a signed commit.

There's also the much simpler pitfall of an attacker just creating a branch named the same as a commit, in the hopes that people will accidentally check it out instead.

Then use git in sha2 mode. You just have to turn that on.

For Google Play: https://developer.android.com/guide/app-bundle/code-transpar...

As far as I know there's no centralised log, it's left up to app developers to publish their key/a log of transparency files.

Using a build service like that is apriori saying "i'm not valuable enough for a targeted attack".

It's easy to defeat right now because very few are currently thinking about secure build systems.

As an example, systemd's latest source tarball has two Web Open Font Format (WOFF) files in a documentation folder, a favicon.png, two few small screenshots and error messages that have unexplained 128bit identifiers. There are also translated strings (PO files) which could include obscure writing systems few can quickly audit, and thus could be advantageous to an attacker wanting to hide malicious code.

The problem with most build systems is the entire source tarball is extracted into a single location and any build scripts executed during build have access to everything in the source tarball. Gentoo's portage does apply sandboxing around an entire build process, just not internally between the different build stages.

Continuing the Gentoo example (being one of the better examples of sandboxed builds), ideally src_unpack could take a compressed tarball distributed from an upstream project and split files into multiple separate paths such as:

- source code files and build scripts needed to build binaries

- supporting test files needed to test the the built binaries

- supporting documentation/binary/high entropy data such as application icons, game data, documentation, etc that should be copied upon installation but aren't required to build binaries

Then src_prepare, src_configure, src_compile, src_test and src_install all have different sandbox configurations restricting which types of files separated during src_unpack each build phase can interact with.

For the systemd example above, some possible improvements could be:

1. Remove WOFF font files and use system default fonts, omit favicon.png and omit screenshots. Or only make WOFF font files, favicon.png and screenshots available for copying during the src_install phase, and ensure they are not readable by build scripts during src_configure and src_compile.

2. Generate error message identifiers using an explained approach such as hash_algorithm("systemd_error_" + name_of_error_constant) to generate nothing-up-my-sleeve identifiers.

3. Only provide access to and include in the build process any translation files of languages the user cares about. Or only make translation files available for copying during the src_install phase and possibly src_test too, and ensure they are not readable by build scripts.

These build system security techniques are obviously more work, but are generally straightforward to understand and implement. These technqiues are in the realm of possibility for smaller embedded Linux systems that may just be kernel + BusyBox + a few small scripts + a bespoke application + a handful of dynamic libraries. And for more complex Linux systems, these techniques are within the realm of possibility when targeted towards high value software, such as software requiring or often executed with root permissions and software requiring simultaneous access to the Internet and access to user files.

Yeah I remember Google's certificate transparency team basically designing firmware transparency for all of Linux (not just Android) as well.