CN102096780B - Rapid detection method of digital fingerprints under large-scale user environment - Google Patents

Abstract

The invention provides a fast detection method of a plurality of digital fingerprints under a large-scale user environment. The method specifically comprises the steps: building a digital fingerprint hash table, matching the fingerprints extracted from works in the digital fingerprint hash table, and obtaining a fingerprint neighbor set which is similar to the extracted fingerprints in a matching way, wherein a user which corresponds to the digital fingerprint which has the minimum error relative to the extracted fingerprints in the fingerprint neighbor set is an illegal user. An LSH (laser shock hardening)-based fast neighbor searching method adopted by the method can fast answer an inquiry request of a digital fingerprint system and can meet the real-time inquiry requirement of the system. Particularly, when the user amount is huge, compared with the existing detection algorithm, the algorithm is much better in the time efficiency.

Description

The method for quick of digital finger-print under the large-scale consumer environment

Technical field

The invention belongs to field of information security technology, be specifically related to the method for quick of digital finger-print under a kind of large-scale consumer environment.

Background technology

Along with the develop rapidly of computer network and multimedia technology, the digital media applications on the internet is being explosion type to be increased.Increasing media work (image, text, audio frequency, video etc.) is propagated with the form of electronics, so that the storage of these works, copy with propagation and become very convenient quick.Digital product progressively incorporates consumer's life, and this has brought new opportunity for numerous creators and publisher, but also is very easy to cause illegal copies and illegal propagation of copyright simultaneously.

Digitized audio frequency and video were play recording arrangement and were come into huge numbers of families along with China's Digital Television popularizes while, and bootlegging and abuse that Digital Television is online will be more serious.Therefore, how to confirm by the identity to illegal distribution person, and it is charged and disciplines as a warning, and then form a kind of deterrent power of hitting illegal infringement, become the copyright protection problem demanding prompt solution.The copyright protection of digital product problem has become an extremely important and urgent subject under discussion of digital world, is the major obstacle that hinders the information digitalization development.How to prevent digital product by bootlegging and propagation, protection copyright owner's legitimate rights and interests have become and have entered the informationized society urgent problem.

These solution of problems require to implement follow-up mechanism in copyright protection, namely can be supervised and control sale, use, circulation and the storage behavior of copyright copy.Many encryption technologies and digital copyright management (DRM) framework adopts encrypts the copyright of protecting Digital Media end to end.In case but the encrypted media data decrypted after, this protection mechanism is just no longer valid.Digital watermark method then can be used for providing further protection to the content of multimedia after the deciphering.As a branch of digital watermarking, digital finger-print is exactly a kind of copyright tracking technique that solves this class problem, is the method for effectively a kind of and tool potentiality, has become the focus of research.Digital finger-print is with different significant cognizance code---fingerprints, utilizes digital watermark technology to be embedded in the Digital Media, and the Digital Media that then will be embedded with fingerprint is distributed to the user.After the publisher finds copy right piracy, just can determine the source of bootlegging by extracting the fingerprint in the pirate product, the bootlegger is prosecuted, thus the effect of the copyright protection of playing and deterrence.At present, the research in digital finger-print field mainly concentrates on the digital finger-print encoding scheme of the anti-collusion attack of design, and has ignored the test problems, particularly the fast detecting problem of digital finger-print when customer volume sharply increases of digital finger-print.Also do not occur at present the fast detecting article about digital finger-print both at home and abroad, but information retrieval field and Data Mining are very burning hot about the research of Fast Similarity Retrieval (fast similarity search) and nearest neighbor search (nearest-neighbor retrieve).How these correlation techniques are applied to solve the fast detecting problem of digital finger-print, both at home and abroad also rare expert, scholar this problem of touching upon.

Present stage, the detection method of digital finger-print still adopted the detecting algorithm method, and mainly contain two kinds: a kind of is the method for linear dependence, and this method is similar to the matched filtering in the communication; A kind of normalization correlation detection method is namely calculated the cosine value of angle between two vectors.Although although the method for the linear dependence of being familiar with of great use and calculate simply, also has obvious shortcoming.The problem of its maximum is that detected value depends on the vector magnitude of extracting to a great extent from works.This shows this detection method for some simple processing, such as the brightness of image or the volume of reduction music, does not possess robustness.This point can not satisfy the digital finger-print application request, is replaced by the normalization correlation method.The computing method that normalization is relevant are as follows:

$T_N\left(i\right)=\frac{z^T{u}_i}{\sqrt{{\left|\right|z\left|\right|}^2}\text{\·}\sqrt{{\left|\right|u_i\left|\right|}^2}},i=\mathrm{1,2},...,L---\left(1\right)$

Wherein z represents the fingerprint sequence that extracts, u _iRepresent the fingerprint sequence that the i bar is original, L represents the number of all digital finger-prints, has the T of maximum related value in the threshold values scope _N(i) corresponding user namely illegally used user's (illegal use refers to that works are carried out information processing, geometric attack or collusion attack etc. and processes rear illegal the propagation) of works:

$j_{\mathrm{sc}}=\arg \underset{j\∈1,L]}{\max }{T}_N\left(j\right)---\left(2\right)$

Can find out from formula (1), the dimension of the time complexity of normalization relevance algorithms and customer volume L and fingerprint code is linear.Flourish along with network, the network user measures sharply to increase with the network user demand of Digital Media is increased, and the scale of data fingerprint is also inevitable to be increased thereupon.In fact, along with the increase of customer volume, and the raising of the requirement of anti-collusion performance, the code length of digital finger-print can increase inevitably thereupon, and this reduces the tracking efficient of digital finger-print to a great extent.When carrying out the normalization correlativity simultaneously, need a large amount of original figure fingerprints is read in the internal memory, this performance to computing machine also is a greatly challenge.

In the existing digital fingerprint system, the code length of the digital finger-print of the overwhelming majority has all reached the real number of hundreds of position even several kilobits.The dimension of the sum of digital finger-print and digital finger-print is linear in the method and system of normalization correlation detection.The dimension of digital finger-print is longer, and customer volume is larger, and the efficient of normalization correlation detection will be lower.Although the normalization correlation detection is accurately very high, the time that tracks out correct disabled user head and shoulders above the patient degree of user, so must in the situation that does not reduce accuracy rate as far as possible, find out a kind of fast matching algorithm.

Summary of the invention

The object of the present invention is to provide a kind ofly based on digital finger-print method for quick under the large-scale consumer environment of local sensitivity Hash (be called for short LSH, lower with), detection speed is fast, and has higher accuracy rate.

The method for quick of digital finger-print under the large-scale consumer environment, be specially: make up the digital finger-print Hash table, the fingerprint that will extract from works mates in the digital finger-print Hash table, the fingerprint neighbour that obtains and take the fingerprint similar collection, it is the disabled user that the fingerprint neighbour concentrates the user corresponding with the digital finger-print that differs minimum of taking the fingerprint.

As optimal way, described digital finger-print Hash table comprises more than one Hash bucket, uses hash function separate more than in the Hash bucket, and the hash function of any two Hash buckets is entirely not identical.

As optimal way, digital finger-print is carried out making up the digital finger-print Hash table after the normalized again, the fingerprint that extracts is mated after by the same way as normalized again.

Technique effect of the present invention is embodied in: according to the similarity between the digital finger-print of finger-print codes generation, the thought of utilizing Hash to carry out quick neighbor search is introduced in the detection of digital finger-print.Quick neighbor search method based on LSH of the present invention, energy quick-responding digital system of fingerprints query requests can satisfy the requirement of system's real-time query.Especially when customer volume was huge, this algorithm is compared existing detection algorithm in time efficiency huge advantage.The method is carried out Hash to high dimensional data, raw data is transformed into the space of low-dimensional, and high dimensional data is carried out dimensionality reduction.The new code word that transforms is high compression, therefore can be loaded in the internal memory at an easy rate, and new code word size is shorter, can calculate fast.

Description of drawings

Fig. 1 is that digital finger-print of the present invention detects schematic diagram.

Fig. 2 is the Establishing process figure of digital finger-print Hash table.

Fig. 3 is that digital finger-print detects particular flow sheet.

Fig. 4 is test pattern, and wherein 4 (a) figure is " Li Na " image; 4 (b) figure is " capsicum " image; 4 (c) figure is " fishing boat " image; 4 (d) figure is " orangutan " image; 4 (e) are " bridge "; 4 (f) are " aircraft "; 4 (g) are " man and wife "; 4 (h) are " wrist-watch ".

Fig. 5 is comparison diagram working time that the present invention is based on LSH detection method and normalization related detecting method, and wherein Fig. 5 (a) is original normalization correlation detection formula and the present invention is based on the contrast of working time of LSH detection method; Fig. 5 (b) for to behind the correlation detection formula optimization with the present invention is based on the contrast of the working time of LSH detection method.

Fig. 6 the present invention is based on LSH detection method antinoise signal to process the ability of attacking the detection design sketch of the fingerprint collection that extracts behind the image plus noise (NOISE) of 6 (a) for the embedding fingerprint; 6 (b) detect design sketch for the fingerprint collection that the rear picture of the image JPEG compression (JPEG) that embeds fingerprint extracts; 6 (c) are the detection design sketch to the fingerprint collection that adds the generation of Gauss's noise on the original fingerprint.

Fig. 7 is the ability schematic diagram that the present invention is based on LSH detection method resist geometric attacks, and 7 (a) are convergent-divergent (RESC) attack detecting design sketch; 7 (b) are rotation (ROT) attack detecting design sketch; 7 (c) add cutting (ROTCROP) attack detecting design sketch for rotation; 7 (d) add stretching (ROTSCALE) attack detecting design sketch for rotation.

Fig. 8 is the ability schematic diagram that the present invention is based on the anti-collusion attack of LSH method.

Embodiment

Below in conjunction with the drawings and specific embodiments the present invention is described in further details.

Digital finger-print among the present invention refers to some the significant identification codes by the finger-print codes generation, rather than the media fingerprints that adopts in the passive regulation technique, and it is corresponding with the media piece that user and this user buy, and is applied to initiatively follow the trail of the disabled user.

As shown in Figure 1, digital finger-print system of the present invention is the same with general digital finger-print body, comprises two parts: the one, for embed fingerprint and the copy distribution system to distributing with the fingerprint copy to copy; Another part is the tracking system that realization is followed the tracks of and tried illegal distribution person.Wherein the information of user j is provided by the user or passes through a series of mutual rear generations by itself and publisher (or the third party who trusts).It generally includes the descriptor of user's identity information and this time purchasing process.The information of relevant user j will be encoded according to certain rule and will be embedded in the former copy that the publisher will sell.The user directly obtains will providing to the user with the copy of fingerprint with the copy of its fingerprint or by the publisher, and the publisher obtains relevant transaction record with the user simultaneously.Dishonest user may directly distribute his resulting copy or copy processed after again row distribution, also may unite with other users and carry out obtaining after the collusion attack the pseudo-distribution after this of new copy.No matter be which kind of situation, all can stay the finger print information that participates in the unlawful activities user in the copy of illegal distribution.In case the publisher has found illegal copies, he will use corresponding fingerprint extraction and fingerprint decoding technique, and use track algorithm to follow the tracks of illegal distribution person.

And digital finger-print system of the present invention is revised in general digital finger-print system.Digital finger-print is carried out Hash, the code word behind the storage Hash.The hash method that adopts can have a variety of selections, such as study Hash, spectrum Hash, semantic Hash etc.Adopted classical local sensitivity Hash (LSH) method among the present invention.Use the method for Hash, although increased the space of storage Hash table, and may bring the reduction of the accuracy of detection, but promote greatly our detection speed, the stand-by period when having reduced the works publisher and inquiring about the own works that have and whether illegally used.

The LSH method that adopts among the present invention comes from " Locality-Sensitive Hashing Scheme Based on p-StableDistributions " method that SoGG meeting that Mayur Datar equals to hold in New York in September, 2004 proposes.The author proposes in article, and the method for LSH can find accurate neighbour at O (logn) for satisfying the data of determining boundary condition in the time, and wherein n is the size of data set.The core concept of this algorithm is the data point of coming mapping (enum) data to concentrate with some hash functions, is higher than the probability that non-similar point clashes far away with the probability of guaranteeing to clash between the similitude, is described as follows:

Mark is defined as follows in the algorithm:

1.

Represent l _pD dimension space under the norm

2. for arbitrarily

Some v in the space, ‖ v ‖ _pThe l of representation vector v _pNorm.

Data set Q is

A finite subset in the space, its length | Q| is n.

3.B (v, r) representative is centered by data point v, r is the ball of radius.If v ∈ B (q, R) claims that then v is the R-neighbour of q, namely v is similar to q.

LSH family's family For any q, function

It is the strictly decreasing function about t.The in other words increase of the distance of random q and v, the possibility that they clash reduces thereupon, and namely similar q and probability that v clashes are far longer than the probability that non-similar q and v clash.

Three some q arbitrarily, v, u, and satisfy v ∈ B (q, R),

P (‖ q-v ‖)＞p (‖ q-u ‖) is therefore arranged.We can be hashing onto the point among the Q among the lower dimensional space U intuitively.When then inquiring about q, only need to calculate the Hash of q, and only need consider the Neighbor Points of q, the point that namely clashes with q.

In order to reach the working time of demand, we enlarge [0, R] and (R, the gap of+probability that ∞) clashes in the scope.In order to reach this purpose, we unite the several functions of use

Especially after specifying K, define a G={g:S → U of function family _K), make g (v)=(h ₁(v) ..., h _K(v)), h wherein _i∈ H, i=1,2 ..., K.For an integer L, we select L function g independently or equiprobably from G ₁..., g _LAt pretreatment stage, algorithm all deposits the some v among the Q in a barrel g _j(v), (j=1 ..., L).

The below specifies the Hash table of setting up the digital finger-print collection and the implementation step of fingerprint detection.

(1) fingerprint initialization

As shown in Figure 2, at initial phase, at first the digital finger-print collection of integral body carried out normalized, and initiation parameter.Carry out the benefit of normalized, the one, can guarantee that the numerical quantity that participates in calculating is more or less the same, the truncation error in the Avoids or reduces computation process, the 2nd, can guarantee that the distance between the code word behind the Hash is suitable with the distance between the original fingerprint.

(2) make up the digital finger-print Hash table

Then set up the digital finger-print Hash table after the normalized.Made up the individual Hash bucket of L (L is more than or equal to 1) in the Hash table, K hash function of a Hash bucket storage is about to the digital finger-print dimensionality reduction and ties up to K (K is more than or equal to 1).K impact inquiry accuracy rate, L affects query time, and concrete size is according to the practical situations adjustment.Each hash function of storage is separate in the single Hash bucket, and the hash function between the Hash bucket is entirely not identical.Publisher's (or trust third party) deposits the cryptographic hash of all original fingerprints in the Hash table relevant position according to above-mentioned Hash table form, and the linking relationship between the index of the cryptographic hash of record original fingerprint and original fingerprint.

(3) fingerprint detection

The testing process of fingerprint is as shown in Figure 3: the starting stage in the detection of fingerprint also needs query fingerprints is carried out normalized, and itself and the process of setting up the Hash table of digital finger-print collection are adapted.In actual applications, the digital finger-print collection that extracts may be subjected to attack, and in the process of the transmission of word works or in the process of image, coding and decoding video, also can be subject to noise pollution, so that exist very large gap between the fingerprint that extracts and the original fingerprint.By normalized, the fingerprint of original figure fingerprint collection and extraction all is evenly distributed between (1 ,+1), dwindle gap between the two, improve the precision that detects.

According to top description to the hash table algorithm of setting up the fingerprint collection, definition is q through the query fingerprints after the normalized.In order to inquire about q, algorithm is from all bucket g ₁(q) ..., g _L(q) search in.At first use the cryptographic hash of the hash function calculating q of storage in each barrel, search this cryptographic hash at Hash table again, find fingerprint v corresponding to this cryptographic hash by itself and the linking relationship of fingerprint index, fingerprint v is the point that clashes with q.For each the fingerprint v that finds in the bucket, calculate the distance between q and the v, if ‖ q-v is ‖≤R, claim that then q is the R-neighbour of v, R is that empirical value can be by repeatedly test be definite, and the simplest mode is that R is made as infinity, and then all fingerprints that clash all are neighbour's collection corresponding to query point q.All ‖ q-v ‖ are sorted, and fingerprint corresponding to minor increment is accurate neighbour's fingerprint of q, and its corresponding user is the disabled user.

Experimental result

The experimental situation configuration that this experiment is adopted is as follows:

Test experiments adopts eight width of cloth test patterns as shown in Figure 4 among the present invention: " Li Na ", " capsicum ", " fishing boat ", " orangutan ", " bridge ", " aircraft ", " man and wife " and " wrist-watch ".In our experiment, the digital finger-print that adopts is 4544 real numbers of tieing up that generate with the algorithm that Wu Min, He Shan etc. propose in document " Collusion-resistant video fingerprintingfor large user group ".Employed fingerprint collection all is the part fingerprint that this fingerprint is concentrated to be processed obtain in several groups of experiments in back.In the experiment, at first the every width of cloth image in 8 width of cloth images shown in Figure 4 is embedded respectively the image that different fingerprints is embedded with fingerprint in a large number, then with digital finger-print attack software StirMark Benchmark 4 to the image of these fingerprints that are embedded with do signal process attack or geometric attack after obtaining attacking pseudo-this, then pseudo-extract fingerprint this from these, that uses in namely obtaining testing was processed by signal to attack and the query fingerprints collection of geometric attack.

Test experiments is tested and interpretation of result time efficiency and accuracy rate two aspects of fingerprint detection, such as Fig. 5, Fig. 6 and shown in Figure 7.Detection effect in the experiment depends on to a great extent digital finger-print and embeds the robustness of algorithm and fingerprint that the fingerprint detection algorithm extracts and the fidelity between the original fingerprint.

Fig. 5 has illustrated normalization correlation method and contrast based on the time efficiency of the detection method of LSH.In the experiment, because matlab software memory restriction and digital finger-print collection dimension are too large, normalization correlation detection method and based on the contrast of the time efficiency of the detection method of LSH when the present invention has tested customer volume from 2000 users to 20000 user.The present invention is based in the method for LSH and used 10 Hash buckets, each Hash bucket inner face has used 32 different hash functions, is about to the original fingerprint collection and reduces to 32 dimensions.Fig. 5 (a) has shown original normalization correlation detection method and the contrast of the detection time based on the LSH method of the present invention.Fig. 5 (b) shown normalization correlation detection formula (1) is optimized after and the contrast based on the time efficiency of the detection method of LSH adopted among the present invention.Optimizing process to normalization correlation detection method is as follows: at first to original figure fingerprint u _i, i=1,2 ..., L and query fingerprints collection z have carried out normalization, and formula (1) can be changed into T _N(i)=z ^Tu _i, i=1,2 ..., L, the lifting that can bring detection speed after the optimization.Fig. 5 (a) and Fig. 5 (b) have shown that the present invention than normalization correlation detection method, has huge advantage in time efficiency.

In the experiment of Fig. 6, at first the every width of cloth image shown in Fig. 4 is embedded respectively the image that different fingerprints obtain a large amount of embedding fingerprints, the digital finger-print that then with StirMark Benchmark 4 softwares the image that embeds fingerprint is carried out gaussian filtering, adds Gauss's white noise, extracts after the JPEG compression detects and directly original fingerprint is added the various fingerprint collection that are subject to after the plus noise attack of Gauss's white noise simulation with matlab software and detects.A Filtering Template has only been adopted in the experiment of gaussian filtering, and the average accuracy of detection is 0.885.Fig. 6 (a) has showed the accuracy of using based on LSH method detection JPEG attack.As shown in the figure, the various degree JPEG compression attack of testing in the experiment do not have impact substantially on testing result, have kept very high accuracy.Fig. 6 (b) has showed that detection adds the accuracy of Gauss's attacked by noise.The experimental result that Fig. 6 (c) shows is to use matlab software directly after original fingerprint adds Gauss's white noise, uses the design sketch that detects based on the LSH method.As shown in the figure, the present invention's employing is based on the algorithm of LSH, to the JPEG compression, add Gauss's sound good detection effect is arranged, especially for the JPEG compression, basically can both detect normally.For directly adding Gauss's noise at original fingerprint, good detection effect is also arranged, along with reducing of the signal to noise ratio (S/N ratio) that adds Gauss's noise, the effect of detection can improve thereupon.

Fig. 7 has shown the design sketch that the fingerprint collection that the image that embeds fingerprint carried out extracting behind 4 kinds of geometric attacks with Stirmark software is tested.Geometric attack is the difficult problem in digital watermarking field and digital finger-print field always, and various embedding algorithms are all undesirable to the robustness of geometric attack, so the accuracy rate that detects for geometric attack is not high.As shown in the figure, although the detection method based on LSH used in the present invention is lower slightly for the detection efficiency of 4 kinds of geometric attacks, the accuracy rate that detects is in the acceptable scope.

Fig. 8 illustrated 2-50 people average attack, maximum attack, minimumly attack, intermediate value is attacked, intersect, improve negatively attack, random negative detection effect of attacking in these 8 kinds of collusion attack situations.As shown in the figure, the detection method that the present invention is based on LSH still reaches more than 0.93 for collusion attack, for most attack method can be correct detect the disabled user.

Claims (3)

1. the method for quick of digital finger-print under the large-scale consumer environment is specially:

In the copyright distribution phase, the descriptor generating digital fingerprint according to user profile and purchasing process carries out Hash to digital finger-print, and the Hash codes that generates is deposited in the digital finger-print Hash table;

In tracking system, from works, extract digital finger-print, generate Hash codes according to the hash method in the described copyright distribution phase for it, and mate with the Hash codes in the digital finger-print Hash table, obtain the fingerprint neighbour collection similar to extracting digital finger-print, it is the disabled user that the fingerprint neighbour concentrates user corresponding to digital finger-print who differs minimum with the extraction digital finger-print.

2. the method for quick of digital finger-print under the large-scale consumer environment according to claim 1, it is characterized in that, described digital finger-print Hash table comprises more than one Hash bucket, use hash function separate more than in the Hash bucket, the hash function of any two Hash buckets is entirely not identical.

3. the method for quick of digital finger-print under the large-scale consumer environment according to claim 1 and 2, it is characterized in that, digital finger-print is carried out making up the digital finger-print Hash table after the normalized again, the fingerprint that extracts is mated after by the same way as normalized again.

Images