Hacker News Reader

Comment by whoevercares | Hacker News Reader

I’m saddened by the fact many brightest mind might be working behind this

This is probably not that a tricky system to implement, from this study.

So from description, it seems that they have a 2 layer system, one synchronous one and an asynchronous one. The sync one filters images based on purely the md5 hashes, so basically a lookup-table check. Then the image goes to the async OCR service to have the text extracted, if being decided against the censorship, it will be write back to the front layer hash table.

Indeed, this is not very different from Facebook's auto face recognition/tagging features that has been enabled like forever. Just the volume of the system is pretty significant.

So all you would need is an app that very slightly modifies one bit in every image you send at random, and it'll never be caught by the synchronous system? Does the async update retroactively delete older posted images? And if so, how long would it take for that to happen?

Probably, an app that applies some random distortion to your source image, just like your own recaptcha.

Also Adverisal Patch is a thing, so that could be applied to specifically distorted the system in a human invisible way to misguide the machine learning models.

Whether it will be retroactively deleted, it think there is a possibility. If WeChat periodically scan your image and find your image violating the censorship, it will probably delete it and report to the authority.

Again, some encryption needs to apply in this case. Just like in the old days.

Technical measures like these don't have to be watertight to be effective with the 'right' laws in place, this has been demonstrated by the restrictions on bypassing 'content protection measures' in the DMCA.

Those found out to be using means to bypass image censoring will eventually see their 'social credit score' plummet and be kicked off WeChat and/or the 'net.

They say the async is operating on a timescale of seconds (presumably long enough for the batch OCR jobs to get around to it), so you can bypass it, but that won't do you much good.

Too smart for your own good.

Tell the physics department to open more faculty positions.

They would love to if they had funding for faculty positions.

Moneys still green after all

(Or whatever color RMB is associated with)

Red

It’s a legal directive, not a monetary incentive

Considering the choice of hash function of md5, not so bright I guess..

You're right.

CityLab researchers exploited MD5's weakness to answer questions about the system. While not a real problem in practice, it seems clear MD5 was not an ideal choice.

From the article, the researchers generated forbidden and allowed images with a colliding hash to prove WeChat was using MD5. The allowed image was banned in the future as a result.

However, MD5 collision generation has some constraints. It's very hard make an image collide with a particular known hash, but it's feasible (5 hours with a large GPU) to take two images and modify them until their hashes collide. Practically this means exploitation opportunities are rather limited, but a forced collision being possible at all seems non-ideal for an adversarial use case. There's also the risk that future cryptanalysis will further weaken MD5. Seems clear to me WeChat just should have used something like sha256.

As you pointed out, since this is an adversarial use case, a robust cryptographic hash function is the way to go. [For a non-cryptographic Hash function, SeaHash [1] would be the best choice, which is violently fast!]

BLAKE2b would have been the perfect choice given the adversarial nature, as it much secure and faster than MD5. [2]

MD5 is so broken, it's really poor choice for any use case - cryptographic (fundamentally broken) or not (fundamentally slow).

[1] http://ticki.github.io/blog/seahash-explained/

[2] https://leastauthority.com/blog/BLAKE2-harder-better-faster-...

> Practically this means exploitation opportunities are rather limited, but a forced collision being possible at all seems non-ideal for an adversarial use case. There's also the risk that future cryptanalysis will further weaken MD5. Seems clear to me WeChat just should have used something like sha256.

With a billion people using phones, "good enough" is probably good enough.

Given the scale and scope of the Chinese security apparatus, anyone capable of using a GPU to hash out collisions is probably already known to the state. And the handful of collisions are probably not important enough to worry about -- a stealthy Winnie The Poo image isn't a big deal.

Another constraint is going to be cpu/battery usage, md5 probably has better hardware support in mobile processors as well as being faster to compute than a 256bit hash.

This is not cryptography. MD5 is a very popular choice for hashing static assets.

But this is an adversarial use case. You're not trying to cause md5 hash collisions in your own static assets, but WeChat users might benefit from that. See my other comment above.

Big whoop. The hash collisions of MD5 allows an "attacker" to prevent himself from posting his own image. A simpler and cheaper way to perform the same "attack" would be to just not post the image in the first place as the outcome is the same.

Ignoring MD5/image format-specific collision realities, theoretically an attacker could submit a contraband image that collides with a valid, allowed image they may want to remove.

When action is taken on the first image, the collided image could also be censored.

Not with current technologies. There is a big difference between creating an image that collides with a specific target hash and just making two images that collide when you control both. The former is not currently possible, and the latter is.

That being said, you could probably create a pair of colliding images, give one to a news outlet or something, then later post the second (presumably banned) one. The app would on short notice need to decide between banning neither or banning both.

> That being said, you could probably create a pair of colliding images, give one to a news outlet or something, then later post the second (presumably banned) one. The app would on short notice need to decide between banning neither or banning both.

Yeah they did this - except the contraband was automatically recognised and both images were banned via hash.

What’s wrong with the application of MD5 to this process?

Why is MD5 a poor choice in this context?

See my above comment: https://news.ycombinator.com/item?id=20508786

Wrong, it is digital fingerprint, not for encrypting. It is suffice for most cases to prevent the same image to spread, had it been blacklisted. Not all users are that tech savvy.

Reply on news.ycombinator.com