A camera fingerprint privacy protection method against link attacks and impersonation attacks

Abstract

The application discloses a camera fingerprint privacy protection method against link attacks and fake attacks, and comprises the following steps: obtaining at least three sub-blurring noises through Gaussian distribution sampling, obtaining a picture to be processed, generating corresponding number of sub-blurring noise coefficients and noise masks (a matrix composed of 0 or 1) according to the number of sub-blurring noises, and combining all the sub-blurring noises into one blurring noise through the sub-blurring noise coefficients and the noise masks. The application realizes the premise of not affecting the normal passing of a legal user through a camera fingerprint-based identity authentication system, simultaneously solves two security risks of link attacks and fake attacks, can help the user effectively defend against identity link attacks, is far faster than multiple iteration denoising in speed, realizes stable camera fingerprint fake attack detection, and avoids illegal users from stealing the camera fingerprints of legal users to pass through the identity authentication system.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of camera fingerprint security, and particularly relates to a camera fingerprint privacy protection method against linkage attacks and forgery attacks. BACKGROUND

[0002] There are countless digital images taken, transmitted and saved on the Internet every moment. According to the research proposed by Lukas et al. in Digital camera identification from sensor pattern noise, there is a photo response non-uniformity (PRNU) in the digital images taken by cameras, and a reliable and high-detection-accuracy camera fingerprint can be extracted according to this characteristic. The core of a digital camera is an image sensor, which is usually made of silicon. Due to the defects in the manufacturing process, the density and size of the silicon image sensor at each pixel point will be different, which will result in different sensitivities of the image sensor pixels to light, i.e., different abilities to convert light into electrical signals. Because this characteristic is introduced by the hardware defects in the manufacturing process, the camera fingerprint has the characteristics of uniqueness and invariability. Moreover, it is difficult to physically copy an image sensor with the same fingerprint. By extracting the camera fingerprint in an unknown source digital image and calculating the similarity with the camera fingerprint in a known source digital image, image tracing and device authentication can be achieved.

[0003] An unprotected camera fingerprint is vulnerable to linkage attacks and forgery attacks. Criminals can determine that two accounts come from the same user by the camera fingerprints in the pictures uploaded by the two different platform accounts, and then perform identity linkage attacks. Moreover, criminals can steal the camera fingerprint of a user by taking digital images with the user's device, and then impersonate the user to pass the camera fingerprint-based verification system.

[0004] According to the research by Martin in Cell phone camera ballistics: attacks and countermeasures, since the camera fingerprint is embodied as pattern noise in a digital image, a malicious attacker can steal the digital image containing the camera fingerprint of a legitimate user, and thus obtain the camera fingerprint of the legitimate user. Moreover, the attacker can use the means of injecting the camera fingerprint of the legitimate user to implement a camera fingerprint forgery attack, so as to pass the camera fingerprint-based device authentication.

[0005] The camera fingerprint forgery attack includes a direct injection attack and a substitution attack. In the direct injection attack, an attacker extracts the camera fingerprint K of a victim from the photos leaked by the victim, and then injects the camera fingerprint K into a digital image taken by the attacker, so as to impersonate the victim to pass the camera fingerprint-based device authentication. In the substitution attack, an attacker extracts the camera fingerprint K of a victim from the photos leaked by the victim, and then replaces the camera fingerprint K of the attacker with the camera fingerprint K of the victim in the digital image taken by the attacker, so as to impersonate the victim to pass the camera fingerprint-based device authentication.v By adjusting the strength coefficient of the injected fingerprint, the fingerprint can be directly injected into a photo taken by an attacker's camera, so that the similarity between the fingerprint extracted from the forged fingerprint image and the reference fingerprint is sufficient to exceed a threshold.

[0006] I′=(1+αK v )I=(1+αK v (1+K) A )I0

[0007] In fingerprint substitution attacks, the image is first denoised to remove as much of the original camera fingerprint as possible, and then fingerprint injection is performed.

[0008] I0′=F(I)

[0009] I′=(1+αK v )I0′=(1+αK v (I)

[0010] Currently, against link attacks, the only way to remove noise from the image is through multiple iterations of denoising. However, this method is time-consuming and cannot effectively prevent camera fingerprint leakage. Regarding camera fingerprint spoofing attacks, Goljan proposed a triangulation detection scheme in "Defending against fingerprint-copy attack in sensor-based camera identification" and Quiring's fragile fingerprint defense scheme in "Fragile sensor fingerprint camera identification." The former requires knowledge of the source of the stolen camera fingerprint, while the latter requires the image from which the attacker steals the fingerprint to be compressed using JPEG. Neither scheme can effectively and stably defend against spoofing attacks.

[0011] In summary, current privacy protection solutions targeting camera fingerprint linking and spoofing attacks still face some technical challenges.

[0012] 1. The method of avoiding link attacks by denoising through multiple iterations is time-consuming and cannot effectively prevent camera fingerprint leakage.

[0013] 2. Forgery attack defense solutions such as triangulation detection and vulnerable fingerprint detection have strong prerequisites and cannot be applied to most situations.

[0014] 3. Currently, there is a lack of technology that can defend against both link attacks and spoofing attacks. Summary of the Invention

[0015] The present application improves the prior art, and provides a camera fingerprint privacy protection method against link attacks and forgery attacks, in particular, the present application is realized through the following technical solutions:

[0016] The present application discloses a camera fingerprint privacy protection method against link attacks and forgery attacks, comprising:

[0017] At least three sub-blurring noises are obtained by Gaussian distribution sampling;

[0018] An image to be blurred is obtained;

[0019] A corresponding number of sub-blurring noise coefficients and noise masks (a matrix composed of 0 or 1) are generated according to the number of sub-blurring noises, and all the sub-blurring noises are combined into one blurring noise through the sub-blurring noise coefficients and the noise masks;

[0020] A Hadamard product is calculated through the randomly generated Gaussian noise A and the blurring noise obtained in the previous step, a slight perturbation is added to the blurring noise, and the blurring noise after the slight perturbation is embedded in the image to be blurred to obtain a blurred image, which is used for uploading, sharing or saving in the future;

[0021] An image to be authenticated is obtained;

[0022] A camera fingerprint in the image to be authenticated is extracted through a camera fingerprint extraction algorithm;

[0023] PCEs are calculated between the camera fingerprint and all the sub-blurring noises, and whether the camera fingerprint in the image to be authenticated is blurred is judged according to all the calculated PCE results;

[0024] The image after the blurring is not authenticated, and the camera fingerprint of the image not after the blurring is matched with the reference camera fingerprint of the user in the database, if the matching is successful, the identity authentication is successfully passed, and if the matching is unsuccessful, the authentication is not passed.

[0025] As a further improvement, the present application generates a corresponding number of sub-blurring noise coefficients and noise masks (a matrix composed of 0 or 1) according to the number of sub-blurring noises, and combines multiple sub-blurring noises into a blurring noise through the sub-blurring noise coefficients and the noise masks, to avoid criminals using enough images after the same blurring to extract the blurring noise and remove it, in particular:

[0026] According to the size of the input image, a noise mask of a corresponding size (a matrix with each data being 0 or 1) is generated, and the sum of all the noise masks is a full 1 matrix:

[0027] ∑B i =E

[0028] The plurality of sub-blurring noises are combined using different noise masks B and sub-blurring noise coefficient combinations, and the sub-blurring noise coefficient is a floating point number from 0 to 1;

[0029] O A = ∑ α i B i O i

[0030] In each blurring process, the same sub-blurring noise, one-time noise mask B and sub-blurring noise coefficient are used to generate different blurring noises, the size of the blurring fingerprint is the same as that of the original image, and the more sub-blurring fingerprints used to form the blurring fingerprint, the safer the blurring fingerprint, but the difficulty of blurring fingerprint detection is also increased.

[0031] As a further improvement, the present application calculates the Hadamard product of the randomly generated Gaussian noise A and the blurring noise to add a slight disturbance to the blurring noise, and embeds the slightly disturbed blurring noise in the picture to obtain the picture for blurring processing, specifically:

[0032] Generate one-time Gaussian noise A and calculate the Hadamard product of the blurring fingerprint to realize blurring noise disturbance hypothesis A, O A ∈ C m*n And A = [a ij ], Then the Hadamard product between them is:

[0033]

[0034] Then the disturbed blurring noise is embedded in the input image in a superimposed manner to complete the blurring processing of the fingerprint:

[0035] I s = I + AO A .

[0036] As a further improvement, the present application calculates the PCE between the camera fingerprint and all sub-blurring noises, and judges whether the camera fingerprint in the to-be-authenticated picture is blurred according to all the calculated PCE results, specifically:

[0037] The camera fingerprint FP extracted from the to-be-detected photo is calculated with each sub-blurring noise O A of the blurring noise O i to obtain a plurality of PCE values:

[0038] PCE i = PCE(Fp, O i )

[0039] As long as any one of the PCE values is abnormal, it is considered that the picture is subjected to a camera fingerprint forgery attack, and only when all the PCE values are around zero, it is considered that the picture is not subjected to a camera fingerprint forgery attack;

[0040]

[0041] The beneficial effects of the present application are as follows:

[0042] 1. By the design of mutual cooperation of the user end and the detection end, the two security risks of link attack and forgery attack are solved without affecting the normal passing of the legitimate user through the camera fingerprint-based identity authentication system.

[0043] 2. The pictures not used by the user for identity authentication are subjected to blurring processing at the user end, which can help the user effectively defend against identity link attack, and the speed is far faster than multiple iterations of denoising.

[0044] 3. The detection of the blurring noise is added before the traditional camera fingerprint authentication at the detection end. By cooperating with the blurring processing at the user end, stable camera fingerprint forgery attack detection is realized, and the illegal user's camera fingerprint is prevented from being stolen to pass through the identity authentication system.

[0045] 4. By the way of blurring processing through multiple sub-blurring noise combinations, the illegal user is prevented from extracting and removing the blurring noise by using enough photos subjected to the same blurring processing, and the security of the defense scheme is improved.

[0046] The present application realizes a fast, effective and stable camera fingerprint privacy protection technology.

[0047] Fast: no need for multiple filtering operations, only a simple matrix operation is needed. And the picture subjected to fingerprint blurring processing has no visual defects.

[0048] Effective: the pictures used by the user for identity authentication are not processed, and the normal passing of the legitimate user through the identity authentication system is not affected; through the blurring processing, the pictures taken by different devices and subjected to the same sub-blurring noise processing also have a similarity greater than zero, and the identity link attack of the illegal user is successfully defended; through the blurring fingerprint detection, whether the camera fingerprint contains the trace of blurring processing is detected, and once the trace of blurring processing exists, it is considered that the camera fingerprint is stolen from the user privacy protection picture by the illegal user, and the forgery attack can be easily detected.

[0049] Stable: Even if the attacker extracts the blur noise from a certain amount of pictures that are processed by the same method of fingerprint blurring, there is no way to remove the blur noise in the picture. When the intensity of each component of the blur noise extracted by the attacker is inconsistent with the intensity of each component on the picture to be removed, it is impossible to remove all the blur noise components at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a data flow diagram of user-side blurring;

[0051] Figure 2 is a data flow diagram of identity authentication at the detection end. DETAILED DESCRIPTION

[0052] The technical solutions of the present application will be further described below by combining with the drawings in the specification and through specific embodiments:

[0053] The purpose of the present application is to propose a privacy protection method that simultaneously solves the two security risks of camera fingerprint, namely, link attack and forgery attack.

[0054] The present application includes two stages.

[0055] 1. The blur noise injection stage is as shown in Figure 1 , deployed at the user end, and the user saves to the local, shares with others or uploads to the social platform, etc. Digital images that need privacy protection all need to undergo fingerprint blurring. After the digital image is subjected to fingerprint blurring, it can defend against user link attacks and resist theft of camera fingerprints by illegal persons.

[0056] 2. The blur noise detection stage is as shown in Figure 2 , deployed at the server that needs to use the camera fingerprint for device authentication. Before the general camera fingerprint authentication, the detection of blur noise is first performed. The image with blur noise is not passed, and the image without blur noise is subjected to subsequent camera fingerprint authentication.

[0057] The specific implementation method of the present application is as follows:

[0058] Blur noise injection stage:

[0059] Step one, at least three sub-blur noises are sampled from the corresponding Gaussian distribution according to the mean and variance of the camera fingerprint distribution, three in this embodiment. The same sub-blur noise is used for blurring each time. It only needs to be generated and saved when blurring is performed for the first time, and the saved sub-blur noise is read for subsequent blurring.

[0060] Step two, generate a corresponding number of sub-blur noise coefficients randomly between 0 and 1 according to the number of sub-blur noises each time the blur processing is performed. Generate a corresponding number and size of noise masks (matrices composed of 0 or 1) randomly according to the number of sub-blur noises and the size of the image to be processed.

[0061] In order to fully utilize all the information of the sub-blur noise, the noise mask needs to satisfy and be a matrix of all ones:

[0062] ∑B i =E

[0063] Where B i represents the noise mask corresponding to the i-th sub-blur noise, and E represents a matrix of all ones.

[0064] Step three, use the randomly generated sub-blur noise coefficients and noise masks to combine the fixed sub-blur noise O i into different blur noise O A :

[0065] O A =∑a i B i O i

[0066] Where O i represents the i-th sub-blur noise, a i represents the sub-blur noise coefficient corresponding to the i-th sub-blur noise, B i represents the noise mask corresponding to the i-th sub-blur noise, and O A represents the blur noise combined by this blur processing.

[0067] Step four, randomly generate a matrix A with the same size as the image to be processed, and add perturbation to the blur noise by calculating the Hadamard product of matrix A and the blur noise. Assuming A, O A ∈C m*n and A = [a ij ], then their Hadamard product is:

[0068]

[0069] Step five, embed the perturbed blur noise into the image to be processed to complete the blur processing:

[0070] I p =I+AO A

[0071] Where O AA denotes a one-time Gaussian noise used to add perturbation, I denotes a picture to be processed, I p denotes a picture after blur processing.

[0072] Blur noise detection stage:

[0073] Step one: Extract camera fingerprint Fp in the picture by camera fingerprint extraction algorithm.

[0074] Step two: Calculate PCE of the extracted camera fingerprint and all sub-blur fingerprints to obtain multiple PCE results:

[0075] PCE i = PCE(Fp, O i )

[0076] Wherein Fp denotes the camera fingerprint extracted from the picture to be tested, O i denotes the i-th sub-blur noise used for blur processing.

[0077] PCE (peak-to-correlation energy) is a classical algorithm in camera fingerprint similarity calculation, based on cyclic cosine similarity calculation:

[0078]

[0079]

[0080] Wherein K1 and K2 denote two camera fingerprints for calculating PCE. C(i, j) denotes the cosine similarity between the two camera fingerprints when the two coordinate axes are respectively shifted by i and j. Sign denotes a sign function, and C(0, 0) is 1 if it is an integer, and -1 if it is negative.

[0081] Step three: Determine whether the camera fingerprint in the picture is subjected to blur processing by the calculated PCE value. If any one of the PCE values is abnormal, it is considered that the picture is subjected to camera fingerprint forgery attack. Only when all PCE values are near zero, it is considered that the picture is not subjected to camera fingerprint forgery attack.

[0082]

[0083] Step four: Perform traditional camera fingerprint detection on the picture after blur fingerprint detection, directly calculate PCE with the user reference fingerprint saved in the database, and compare with the threshold value:

[0084] PCE = PCE(Fp, F ref )

[0085]

[0086] The above description is not intended to be a restriction on the present application. It is to be understood that numerous variations and modifications can be made by those skilled in the art, which fall within the spirit of the application, and their equivalents, without departing from the scope of the application in its broadest form.

Claims

1. A camera fingerprint privacy protection method against link attacks and forgery attacks, characterized in that, The method comprises the following steps: Sampling at least three sub-blurring noises through Gaussian distribution; Obtaining a picture to be processed; Generating a corresponding number of sub-blurring noise coefficients and noise masks according to the number of sub-blurring noises, wherein the noise mask is a matrix composed of 0 or 1, and combining all the sub-blurring noises into one blurring noise through the sub-blurring noise coefficients and the noise mask; Adding slight perturbation to the blurring noise by calculating the Hadamard product of the randomly generated Gaussian noise A and the blurring noise obtained in the previous step, embedding the slightly perturbed blurring noise in the picture to be processed to obtain a picture processed by blurring, which is used for uploading, sharing or saving in the future; Obtaining a picture to be authenticated; Extracting the camera fingerprint in the picture to be authenticated through a camera fingerprint extraction algorithm; Calculating the PCE between the camera fingerprint and all the sub-blurring noises, and judging whether the camera fingerprint in the picture to be authenticated is processed by blurring according to all the calculated PCE results; The picture processed by blurring is not authenticated, and the camera fingerprint of the picture not processed by blurring is matched with the reference camera fingerprint of the user in the database, and if the matching is successful, the identity authentication is passed, otherwise, the authentication is failed; the step of generating a corresponding number of sub-blurring noise coefficients and noise masks according to the number of sub-blurring noises, and combining multiple sub-blurring noises into a blurring noise through the sub-blurring noise coefficients and the noise mask, is used to avoid that the criminals use enough pictures processed by blurring to extract the blurring noise and remove it, and the specific steps are as follows: Generating a noise mask corresponding to the size of the input picture, and the sum of all the noise masks is a full 1 matrix: ∑B i = E Multiple sub-blurring noises are combined using different noise masks B and sub-blurring noise coefficients, and the sub-blurring noise coefficients are floating-point numbers from 0 to 1; O A =∑α i B i O i In each blurring process, the same sub-blurring noise, one-time noise mask B and sub-blurring noise coefficient are used to generate different blurring noises, the size of the blurring fingerprint is the same as that of the original image, and the more sub-blurring fingerprints used to compose the blurring fingerprint, the safer the blurring fingerprint is, but the difficulty of blurring fingerprint detection is also increased.

2. The camera fingerprint privacy protection method against link attacks and forgery attacks of claim 1, wherein, The specific steps of calculating the Hadamard product of the randomly generated Gaussian noise A and the blurring noise to add slight perturbation to the blurring noise, and embedding the slightly perturbed blurring noise in the picture to obtain a picture processed by blurring are as follows: Generate one-time Gaussian noise A and fuzzy fingerprint calculation Hadamard product to realize fuzzy noise disturbance, assuming A, O A ∈C m*n And A = [a ij ], Then the Hadamard product between them is: Then, the perturbed blurring noise is embedded in the input image in a superimposed manner to complete the blurring of the fingerprint: I p = I + AO A .

3. The camera fingerprint privacy protection method against link attacks and forgery attacks of claim 1, wherein, The specific steps of calculating the PCE between the camera fingerprint and all the sub-blurring noises, and judging whether the camera fingerprint in the picture to be authenticated is processed by blurring according to all the calculated PCE results are as follows: The camera fingerprint FP extracted from the photo to be detected is compared with each sub-fuzzy noise O A composing the fuzzy noise O i Each sub-fuzzy noise O i is subjected to PCE calculation to obtain a plurality of PCE values: PCE i = PCE(Fp, O i ) If any one of the PCE values is abnormal, it is considered that the picture has been subjected to camera fingerprint forgery attack, and only when all the PCE values are near zero, it is considered that the picture has not been subjected to camera fingerprint forgery attack.

Images