Decryption: Ali found the leaker just by taking a screenshot, and I knelt directly after reading the technical analysis

Alibaba found the leaker just by screenshots. How can technology do this?

>>>>@ liu fuqiang's sharing
Through an experiment, this paper briefly introduces the method of adding digital blind watermarking in frequency domain, and further verifies its anti-attack ability. Based on the above experiments, the methods of avoiding digital blind watermarking are summarized. (Multi image warning)
This paper is divided into five parts,
The first part is an overview;
The second part introduces the principle of making blind digital watermarking in frequency domain;
The third part is blind watermark attack experiment;
The fourth part is the summary;
Part V Appendix (source code).
1、 Overview
This article provides a technical means to realize "Alibaba tracks employees through identification codes that cannot be recognized by the naked eye". According to the analysis of other respondents, Ali may not have used the frequency domain watermarking technology.
Compared with the spatial method, the method of adding blind watermark in frequency domain is more stealthy and more resistant to attack. This kind of algorithm is difficult to solve the watermark. You don't know which frequency band the watermark is added to, and the attack will often destroy the original content of the image. In this paper, the popular science adds digital blind watermarking by means of frequency domain. For the web, you can add a background image to track the screenshot.
The so-called blind watermark refers to the watermark that can't be perceived by people, including invisible or inaudible (yes, digital blind watermark can also be used for audio). It is mainly used in audiovisual works, digital books, etc., with the purpose of protecting and tracking copyright without damaging the original works.
The methods of adding digital blind watermarking can be simply divided into spatial domain method and frequency domain method. These two methods add redundant information, but the original image size will not change when the coding and compression conditions remain unchanged (the original is 10MB after adding blind watermarking or 10MB).
The airspace refers to the space domain, and the image we see everyday is the airspace. The method of adding digital watermark in the spatial domain is to directly operate on the image in the spatial domain (the reason for this is that not only the original image is the spatial domain, but also the difference of the original image is the spatial domain), such as directly overlaying the watermark on the image.
We often say how high a tone is. This pitch refers to the frequency; Similarly, the strong change of image gray level can also be regarded as the frequency of the image. The method of adding digital watermark in the frequency domain refers to transforming the image into the frequency domain (wavelet domain) by some transformation means (Fourier transform, discrete cosine transform, wavelet transform, etc.), adding watermark to the image in the frequency domain, and then converting the image into the space domain through inverse transformation. Compared with airspace means, frequency-domain means are more stealthy and more resistant to attack.
The so-called attack on the watermark refers to destroying the watermark, including smearing, cutting, scaling, rotating, compressing, noising, filtering, etc. Digital blind watermarking should not only be agile (not to be caught), but also be defensive (anti attack). Just like Dota's agile heroes are often fragile, the concealment and robustness of digital blind watermarking are mutually exclusive. (Robustness is the academic name of anti-attack)
2、 A method of making digital blind watermark in frequency domain
The signal has a frequency. A signal can be seen as the superposition of countless sine signals of different orders.

The above formula is Fourier transform formula,
It refers to time domain signal (for signal, we say time domain because it is related to time, while for image, we often say space domain and space)
In short, we have a way to convert the time domain signal into the frequency domain. Similarly, we can also convert the two-dimensional signal (image) into the frequency domain. As mentioned above, the frequency of the image refers to the intensity of the image gray transformation. For more systematic knowledge of this aspect, see Gonzalez's "Image Processing".
Take Fourier transform as an example to introduce the method of adding digital blind watermark to the image in the frequency domain. Note that because the image is a discrete signal, we actually use the discrete Fourier transform. In this paper, we use the two-dimensional fast Fourier transform. The fast Fourier transform is equivalent to the discrete time Fourier transform. Through butterfly merging, the speed is faster. Fourier transform in the following is two-dimensional fast Fourier transform.

The above figure shows the basic process of superimposed digital blind watermarking. The purpose of coding is to encrypt the watermark and control the distribution of watermark energy. The following is an experiment of superimposed digital blind watermarking. This is the original image, size 300 * 240

Then Fourier transform is performed, and the frequency domain image after transformation is shown in the figure below

This is the watermark I want to add. The size is 200 * 100

This is my encoded watermark. The encoding method is random sequence encoding. Through encoding, the watermark is distributed to random frequencies, and the watermark is encrypted

Superimposing the spectrum of the above image and the original image shows that the spectrum of the image has changed greatly

Then, the spectrum of the superimposed watermark is inverted Fourier transform to obtain the image after the superimposed digital watermark

The difference between the superimposed image and the original image is almost invisible to the naked eye, so the digital blind watermark has been superimposed into the image

In fact, we add the watermark to the original image in the form of noise

The following figure shows the residual difference between the watermarked map and the original map in the airspace (the contrast is adjusted, otherwise the residual difference will be reduced and invisible)

It can be seen that in fact, the above method is to add redundant information (like noise) through frequency domain. These noises spread all over the map and are not easily damaged in the airspace

Finally, the mean square error (MSE) is 0.0244

Signal to noise ratio (PSNR) is 64.2dB

So why has the spectrum changed so much, but the change in the airspace is so small? This is because we avoid the main frequency of the image. The figure below shows the vertical spectrum of the original image, and its energy is mainly concentrated in low frequency

Watermark extraction is the inverse process of watermark superposition

After extraction, we get the following watermark: Why is the watermark symmetrical

3、 Aggressive experiment

This part carries out attack experiments to verify the robustness of superimposed digital blind watermarking by means of frequency domain

1. Smear attack. This is the image after the attack:

Then extract the watermark:

2. Cutting attack is the situation of taking a part of screenshots often used on the Internet:

Carry out cyclic completion:

Extract watermark:

3. Scaling attack (this experiment is made in plain code, the watermark energy is high, and the concealment is not strong):

Extract watermark (watermark is not good, and frequency mixing is very serious):

4. Rotation attack (clear code):