Reverse engineering censored videos is interesting thing really, I had no idea it's that simple to do:

I think it might be relevant since musicians and producers may do video tutorials showing cracked plugins hiding actual cracks

 

Reverse engineering is the process of gaining sufficient understanding of a product at the design level to support its maintenance, improvement, or replacement. In software development, it is a part of the software maintenance process that helps gain sufficient understanding of the system to make decisions about it. This process does not alter the system in any way.

 

I remember a news story about a notorious pedophile who used a spiral blur on his face. They swirled it the reverse way and caught him.

 

Anything with a pattern beneath it, AI will detect it. That's its power. That's also what makes it scary because human behavior is exactly that: a pattern. That's how predictive policing works: it's basically an AI tool that detects sociological/criminologist patterns and can therefore predict crime. That's just one example, the possibilities are endless. Just as those blurred pixels 'hiding' text, you may think your behavior is hidden, but it is not.

 

Maybe AI is able to decode video clips (in astronomy is used a similar principle: overlap several different pics of a whatever sky object to obtain the highest possible detail), but is very difficult with one single frame, and the result is not plausible.

I remember a movie with Kevin Costner where a computer tried succesfully, though within a long time span, to rebuild a detailed picture of a spy from a single badly blurred photo.
That was SF then and SF now. In future... maybe still SF.

You just can't get info if missing, just like a sampled sound in 8 bit cannot be converted in a 32 bit and pretend it is the same quality of that sound sampled in 32 bits.

 

indeed, the point of the video is, on moving objects (pixelated or blurred), those can be reverse engineered, obviously if you have a picture, it would be impossible

 

I think it is well known by now that you shouldn't use any effect that calculates a result to hide something as someone might be able to undo the calculation. If you use a single color block or a pattern without transparency, you will be safe.

 

AI model for de-pixelation trained on 241 TB of Japanese adult video.

 

LOL neither pixelation, blurring, nor blackening (redaction) are consistently secure methods for data anonymization. With pixelation and blurring, the anonymized data can often be reconstructed sufficently well using modern algorithms to be read - even with a combination of both methods or if noise was previously present, as long as enough data points are available. while this has been possible in principle for some time, it's easier than ever today thanks to AI tools.
Methods like redaction are also often not as secure as many assume.
With digital text redaction, a still widespread mistake is using separate layers that can simply be removed. The indexing of redactions in metadata also significantly aid reconstruction efforts.
However, whether it's digital or analog doesn't matter when it comes to the length of the redacted bars.
With partial redactions, conclusions about suitable text candidates can be drawn from the character count or length of the redacted text sections. If a black block of the same length appears multiple times in the text it may be possible to determine from the context whether it's the same word. If this is likely, the pool of candidates can often be significantly narrowed down.
Anonymization using person and object redaction in images and videos is also vulnerable to a certain degree. Just as with text redaction, although the pixel data itself is irretreivably removed, this doesn't mean that information can no longer be derived from the remaining data. shadows, reflections, movements, the silhouette of the redaction, the position and posture of the object or person, and the surroundings – all of these are pieces of information from which a more accurate picture of the hidden information can often be derived.
Physical masking, such as face masks or scarves, also often offers only limited protection against de-anonymization: biometric analysis algorithms can often identify individuals simply based on the remaining visible facial parts (cheekbones, eyes, forehead, etc), especially if the covering fits snugly and the basic facial shape remains recognizable. So, sunglasses and a medical mask offer relatively little help if you don't want to be recognized. Only voluminous coverings that significantly alter the silhouette still pose a greater hurdle, at least for now (as far as I know).
Voice alteration, especially before the era of AI voice cloning/changing, also often failed to provide secure protection. Even voices heavily altered by complex algorithm chains (combinations of pitch shifting, vocoder, distortion, frequency modulation, granular synthesis, etc etc) can be reconstructed through audio forensics. While perfect reconstruction is almost never possible, it's usually feasible to restore the voice to the point where it becomes recognizable and atributable for people who know the original. Well, even if the altered voice cannot be sufficiently reconstructed, conclusions about identity can still be drawn from timing, rhythm, frequency patterns, etc. And to what extent AI voice changing can be reconstructed, I honestly don't know. In any case, I wouldn't rely on it being impossible.
And anyone who believes they are a ghost on the internet just because they use a vpn and email forwarding is often simply naive. Most of us leave such a broad trail of digital breadcrumbs that, with some effort and the right tools, we can often be identified with sufficient certainty.

 

No, it's not impossible. There is less data, but even within this data set there are matches from which parts of the information can be reconstructed

 

So silly hahahah.



ermmm what's the program called?

 

Yes, modern tools can reconstruct a lot of detail but the result is rarely the ground truth.
Works better for known shapes and worse for the unknown.

 

BOOM !!

 

It depends on what the model was trained with and how many clues there are in the image. Let's take the example from the video:
It's obvious that it's macOS. If there was sufficient macOS UI material in the training data, the top bar and sidebar on the finder window are known patterns.
This might already be sufficient to reconstruct a lot of the unknown, because it was pixelated by the same algorithm.
Edit: Of course, this is an estimate based on probabilities (as you said), but it could be more than enough to get some interesting insights and test some vulnerabilities.

 

Last edited: Apr 17, 2025

I'm sure it's great for unpixelating faces.

 

Last edited: Apr 17, 2025

Very disturbing

Mr Swirl: The Internet's Most Disturbed User

 

At first read I thought that this thread was about videos concerning reverse engineering that are censored on yutube. I'm a wee disappointed.

 

Here I am fulfilling @Lois Lane's record wish!

Queen - Another One Bites The Dust - Backwards