Cryptographic obfuscation and ‘unhackable’ software

I have a thing for over-the-top cryptography headlines — mostly because I enjoy watching steam come out of researchers’ ears when their work gets totally misrepresented. And although I’ve seen quite a few good ones, last week WIRED managed a doozy.

The headline in question, Cryptography Breakthrough Could Make Software Unhackable, managed to accomplish something that few cryptography headlines do. It sent its own protagonist, Amit Sahai, into the comments section to perform intellectual garbage pickup.

In contrast to the headline, which is quite bad, the article is actually pretty decent. Still, the discussion around it has definitely led to some confusion. As a result, many now think an amazing breakthrough has taken place — one that will finally make software secure. They’re probably right about the first part. They may be disappointed about the rest.

The truth, as usual, is complicated. There is, indeed, something very neat going on with the new obfuscation results. They’re just not likely to make software ‘unhackable’ anytime soon. They might, however, radically expand what we can do with cryptography. Someday. When they’re ready. And in ways we don’t fully understand yet.

But before I go into all that, it’s probably helpful to give some background.

Program obfuscation

The Wired article deals with the subject of ‘program obfuscation‘, which is a term that software developers and cryptographers have long been interested in. The motivation here is pretty simple: find a way that we can give people programs they can run — without letting them figure out how the programs work.

Note that the last part necessarily covers a lot of ground. In principle it includes aspects ranging from the nature of specific secret algorithms used — which may be proprietary and thus worth money — to secret information like passwords and cryptographic keys that might be hardcoded into the program.

For a simple example, consider the following routine:

// Check a password and print out top secret information if it’s correct
//
SuperSecretPasswordProtectedStuff(string passwd) {
  if (password == “0xt438fh27266629zn28366492923aai3jnqobbyc4t!”) {
    print(“Congratulations. Here’s some super secret private information: ….\n”);
  } else {
    print(“Wrong password, fool.\n”);
  }
}

If you’re like me, you probably wrote a program like this at some point in your life. You may have eventually realized how ineffective it would be — against a smart attacker who could dump the program code. This is because most programs (and even compiled binaries) are pretty easy to read. An attacker could just look at this program to recover the secret password.

Program obfuscation is motivated by the idea that many useful programs would benefit if we could somehow ‘stop’ people from doing this, while still letting them possess and run the code on their own computers.

In real world software systems, ‘obfuscation’ usually refers to a collection of ad-hoc techniques that turn nice, sensible programs into a morass of GOTOs and spaghetti code. Sometimes important constants are chopped up and distributed around the code. Some portions of the code may even be encrypted — though only temporarily, since decryption keys must ship with the program so it can actually be run. Malware authors and DRM folks love this kind of obfuscation.

A chunk of ‘birken’, one of the winning entries in the 2013 obfuscated C contest.

While these techniques are nice, they’re not really ‘unhackable’ or even secure against reverse-engineering in the strong sense we’d like. Given enough time, brains and tools, you can get past most common software obfuscation techniques. If you have, say, Charlie Miller or Dion Blazakis on your side, you probably do it quickly. This isn’t just because those guys are smart (though they are), it’s because the techniques are not mathematically rigorous. Most practical obfuscation amounts to roadblocks — designed to slow reverse-engineers down and make them give up.

So what does it mean to securely ‘obfuscate’ a program?

The poor quality of existing software obfuscation set cryptographers up with a neat problem. Specifically, they asked, could we define a strong definition of program obfuscation that would improve on, to put it politely, the crap people were actually using? Given such an obfuscator I could hand you my obfuscated program to run while provably protecting all partial information — except for the legitimate inputs and outputs.

If this could be done, it would have an incredible number of applications. Stock traders could obfuscate their proprietary trading algorithms and send them to the cloud or to end-customers. The resulting programs would still work — in the sense that they produced the right results — but customers would never learn anything about ‘how they worked’. The internals of the program would be secret sauce.

Unfortunately, when the first researchers started looking at this problem, they ran into a very serious problem: nobody had any idea what it meant to securely ‘obfuscate a program in the first place.

Do think about it for a second before you roll your eyes. What do you think it means to obfuscate a program? You’re probably thinking something like ‘people shouldn’t learn stuff‘ about the program. But can you explain what stuff? And does ‘stuff‘ depend on the program? What about programs where you can efficiently learn the whole program from sending it inputs and seeing the results? What does it mean to obfuscate those?

Clearly before progress could begin on solving the problem, cryptographers needed to sit down and figure out what they were trying to do. And indeed, several cryptographers immediately began to do exactly this.

Black-box cryptographic obfuscation

The first definitions cryptographers came up addressed a very powerful type of obfuscation called ‘virtual black box obfuscation‘. Roughly speaking, it starts from the following intuitive thought experiment.

Imagine you have a program P, as well as some code obfuscation technique you’ve developed. It’s important that the obfuscation be efficient, meaning it doesn’t slow the program down too much. To determine whether the obfuscation is successful, you can conduct the following experiment.

  1. Give Alice a copy of the obfuscated program code Obf(P), in a form that she can look at and run on her own computer.
  2. Take the original (unobfuscated) program P, and seal it in a special computer located inside an armored ‘black box’. Let a user Sam interact with the program by sending it inputs and receiving outputs. But don’t let him access the code.

Roughly speaking, what we want from secure obfuscation is that Alice should learn ‘no more‘ information after seeing the obfuscated program code, than could some corresponding user Sam who interacts with the same program locked up in the black box.

What’s nice here is that this is we have the beginnings of an intuitive definition. In some sense the obfuscation should render the program itself basically unintelligible — Alice should not get any more information from seeing the obfuscated program than Sam could get simply by interacting with its input/output interface. If Sam can ‘learn’ how the program works just by talking to it, even that’s ok. What’s not ok is for Alice to learn more than Sam.

The problem with this intuition is, that of course, it’s hard to formulate. One thing you may have noticed is that the user Alice who views the obfuscated program truly does learn something more than the user Sam, who only interacts with it via the black box. When the experiment is over, the user who got the obfuscated program still has a copy of the obfuscated program. The user who interacted with the black box does not.

To give a practical example, let’s say this program P is a certificate signing program that has a digital signing key hard-coded inside of it — say, Trustwave’s — that will happily attach valid digital signatures on certificates (CSRs) of the user’s devising. Both Alice and Sam can formulate certificate signing requests and send them into the program (either the obfuscated copy or the version in the ‘black box’) and get valid, signed certificates out the other end.

But here’s the thing: when the experiment is over and we shut down Sam’s access to the black box, he won’t be able to sign any more certificates. On the other hand, Alice, who actually who learned the obfuscated program Obf(P), will still have the program! This means she can keep signing certificates on and forevermore. Knowing a program that can sign arbitrary certificates is a pretty darn significant ‘piece’ of information for Alice to have!

Worse, there’s no way the ‘black box’ user Sam can ever learn a similar piece of information — at least not if the signature scheme is any good. If Sam could ask the black box for a bunch of signatures, then somehow learn a program that could make more signatures, this would imply a fundamental weakness in the digital signature scheme. This cuts against a basic property we expect of secure signatures.

Barak et al. proposed a clever way to get around this problem — rather than outputting an arbitrary piece of information, Alice and Sam would be restricted to outputting a single bit at the end of their experiment (i.e., computing a predicate function). This helps to avoid the problems above and seems generally like the “right” definition.

An impossibility result

Having proposed this nice definition, Barak et al. went on to do an irritating thing that cryptographers sometimes do: they proved that even their definition doesn’t work. Specifically, they showed that there exist programs that simply can’t be obfuscated under this definition.

The reason is a little wonky, but I’m going to try to give the flavor for it below.

Imagine that you have two programs A and B, where A is similar to our password program above. That is, it contains a hard-coded, cryptographic secret which we’ll denote by password. When you run A(x), the program checks whether (x == password) and if so, it outputs a second cryptographically strong password, which we’ll cleverly denote by password_two. If you run A on any other input, it doesn’t output anything.

Now imagine the second program B works similarly to the first one. It contains both password and password_two hardcoded within it. A major difference is that B doesn’t take a string as input. It takes another computer program. You feed a program in as input to B, which then:

  1. Executes the given program on input password to get a result r.
  2. If (r == password_two)it outputs a secret bit.

It should be apparent to you that if you run the program B on input the program A — that is, you compute B(A) — it will always output the secret bit.* It doesn’t even matter if either B or A has been obfuscated, since obfuscation doesn’t change the behavior of the programs.

At the same time, Barak et al. pointed out that this trick only works if you actually have code for the program A. If I give you access to a black box that contains the programs A, B and will simply let you query them on chosen inputs, you’re screwed. As long as password and password_two are cryptographically strong, you won’t be able to get A to output anything in any reasonable amount of time, and hence won’t be able to get B to output the secret bit.

What this means is that Alice, who gets the obfuscated copies of both programs, will always learn the secret bit. But if Sam is a reasonable user (that is, he can’t run long enough to brute-force the passwords) he won’t be able to learn the secret bit. Fundamentally, the obfuscated pair of programs always gives more information than black-box access to them.

It remains only to show that the two programs can be combined into one program that still can’t be obfuscated. And this is what Barak et al. did, completing their work and showing the general programs can’t be obfuscated using their definition.

Now you can be forgiven for looking askance at this. You might, for example, point out that the example above only shows that it’s hard to obfuscate a specific and mildly ridiculous program. Or that this program is some kind of weird exception. But in general it’s not. There are many other useful programs that also can’t obfuscate, particularly when you try to use them in real systems. These points were quickly pointed out by Barak et al., and later in other practical settings by Goldwasser and Kalai, among others.

You might also think that obfuscation is plain impossible. But here cryptography strikes again — it’s never quite that simple.

We can obfuscate some things!

Before we completely write off black box obfuscation, let’s take a moment to go back to the password checking program I showed at the beginning of this post.

Here’s a slightly simpler version:

// Check a password and return true if it’s correct, false otherwise
//
bool SuperSecretPasswordProtectedStuff(string passwd) {
  if (password == “0xt438fh27266629zn28366492923aai3jnqobbyc4t!”) {
    return true;
  } else {
    return false;
  }
}

This function is an example of a ‘point function‘: that is, a functions that returns false on most inputs, but returns true at exactly one point. As you can see, there is exactly one point (the correct password) that makes this routine happy.

Point functions are interesting for a two simple reasons: the first is that we use them all the time in real systems — password checking programs being the obvious example. The second is that it turns out we can obfuscate them, at least under some strong assumptions. Even better, we can do it in a way that should be pretty familiar to real system designers.

Let be a secure hash function — we’ll get back to what that means in a second — and consider the following obfuscated password checking routine:

// Check a password and return ‘true’ if it’s correct, ‘false’ otherwise
// But do it all *obfuscated* and stuff
//
bool ObfuscatedSuperSecretPasswordProtectedStuff(string passwd) {

  static string HARDCODED_SALT          = 0x……; // this is a salt value

  static string HARDCODED_PASSWORD_HASH = 0x……; // this value is H(salt + target password)

  // First, hash the input password with the salt
  hashedPass = H(HARDCODED_SALT + passwd);

  if (hashedPass == HARDCODED_PASSWORD_HASH) {
    return true;
  } else {
    return false;
  }
}

Note that our ‘obfuscated’ program no longer stores the plaintext of the password. Instead we store its hash only (and salt), and wrap it inside a program that simply compares this hash to a hash of the user’s input. This should be familiar, since it’s the way you should be storing password files today.**

A number of cryptographers looked at formulations like this and showed the following: if the password is very hard to guess (for example, it’s secret and drawn at random from an exponentially-sized space of passwords) and the hash function is ‘strong’ enough, then the hash-checking program counts as a secure ‘obfuscation’ of the basic password comparison program above it.

The intuition for this is fairly simple: imagine Alice and Sam don’t know the password. If the password is strong, i.e., it’s drawn from a large enough space, then their probability of ever getting the program to output ‘true’ is negligible. If we assume an ideal hash function — in the simplest case, a random oracle — then the hard-coded hash value that Alice learns is basically just a useless random string, and hides all partial input about the real password. This means, in practice, that any general question Alice can answer at the end of the experiment, Sam can answer with about the same probability.***

Finding better definitions

More than a decade after the definition was formulated, there are basically two kinds of results about ‘strong’ virtual-black-box obfuscation. The first set shows that it’s impossible to do it for general programs, and moreover, that many of the interesting functions we want to obfuscate (like some signatures and pseudorandom functions) can’t be obfuscated in this powerful sense.

The second class of results shows that we can black-box obfuscate certain functions, but only very limited ones like, say, point functions and re-encryption. These results are neat, but they’re hardly going to set the world on fire.

What cryptographers have been trying to achieve since then is a different — and necessarily weaker — definition that could capture interesting things we wanted from obfuscating general programs, but without all the nasty impossibility results.

And that, finally, brings us to the advances described in the WIRED article.

Indistinguishability Obfuscation

In shooting down their own proposed definitions, Barak et al. also left some breadcrumbs towards a type of obfuscation that might actually work for general programs. They called their definition “indistinguishability obfuscation” (IO) and roughly speaking it says the following:

Imagine we have two programs C1, C2 — which we’ll describe as similarly-sized circuits — that compute the same function. More concretely, let’s say they have exactly the same input/output behavior, although they may be implemented very differently inside. The definition of indistinguishability obfuscation states that it should be possible to obfuscate the two circuits C1, C2 such that no efficient algorithm will be able to tell the difference between Obf(C1) from Obf(C2). 

While this idea was proposed years ago, nobody actually knew how to build any such thing, and it was left as one of those ‘open problems’ that cryptographers love to tear their hair out over. This remained the case until just last year, when a group of authors from UCLA, UT Austin and IBM Research proposed a ‘candidate construction’ for building such obfuscators, based on the new area of multilinear map-based cryptography.

Another interesting variant of this notion is called extractability obfuscation (EO), which implies not only that you can’t distinguish between Obf(C1) and Obf(C2), but moreover, that if you could distinguish the two, then you could necessarily find an input value on which both C1 and C2 would produce different outputs. Moreover, other work indicates IO and EO give essentially the ‘best possible’ obfuscation you can provide for general programs.

The question you’re probably asking is: so what? What can we do with indistinguishability obfuscation?

And this is where the WIRED article differs substantially from the reality. The truth is that IO will probably bring us major cryptographic and software advances — it’s already been shown to bring about succinct functional encryption for all circuits, as well as advances in deniable encryption. Moreover it can be used to build known forms of encryption such as public-key encryption based on new techniques that differ from what we use today — for example, by obfuscating ‘symmetric’ primitives like pseudorandom functions.****

And if this doesn’t seem quite as exciting as the WIRED article would imply, that’s because we’re still learning what the possibilities are. The new techniques could, for example, allow us to build exciting new types of security systems — or else show us radical new ways to build systems that we already have. Even the latter is very important, in case advances in our field result in ‘breaks’ to our existing constructions.

What IO and EO will probably not do is make programs unhackable, since that implies something a whole lot stronger than either of these techniques currently provide. In fact, it’s not clear what the heck you’d need to make programs unhackable.

And the reason for that is quite simple: even obfuscated software can still suck.

Notes:

Thanks to Susan Hohenberger and Zooko for comments on this post. 

The bit in the Barak et al. proof doesn’t have to be secret.

** With a pretty damn big caveat, which is that in real life (as opposed to cryptographic examples) users pick terrible passwords, which makes your password hashes vulnerable to dictionary attacks. We have a variety of countermeasures to slow down these attacks, including salt and computationally intensive password hashing functions.

*** God I hate footnotes. But it’s worth unpacking this. Imagine that Alice learns ‘some information’ from seeing an average program containing the hash of a strong password. If the password is strong, then Alice can only guess the right input password with negligible probability. Then Sam can ‘use’ Alice to get the same information as follows. He formulates a ‘fake’ program that contains a completely bogus password hash and mails it to Alice. Whatever information she learns from his program, he takes and outputs as the information he ‘learned’. It remains to show that if the password hash is strong (a random oracle), Alice isn’t going to be able to learn more from a random hash than she would from the hash of a strong password.

**** The basic idea here is that symmetric encryption (like AES, say) can’t be used for public key encryption, since anyone who learns your encryption key also learns your decryption key. But if you can obfuscate an encryption program that contains your symmetric encryption key, then the new program becomes like a public key. You can hand it out to people to encrypt with.