A face emotion recognition method for protecting data privacy in mobile crowd sensing

Abstract

The application discloses a kind of mobile group perception data privacy protection face emotion recognition method, by the extraction of feature of inquiring user local computing power and encryption, matching is carried out on cloud by the way of data outsourcing, then the matching condition is obtained to recognize face expression data, it is related to including face feature vector provider, two cloud servers CP, CSP of providing strong computing power and storing encrypted data and multiple mobile devices for face emotion feature recognition.The application method uses homomorphic encryption technology, allows feature vector matching in ciphertext state, thereby improves the computing efficiency and data privacy protection level, reduces the risk of personal privacy leakage, and provides higher privacy protection degree, with high application flexibility, can adapt to the needs of different application scenarios, including medical care, social media analysis and personalized recommendation, etc., is worth promoting.

Description

Technical Field

[0001] This invention relates to the field of information security technology, and in particular to a facial emotion recognition method for data privacy protection in mobile group perception. Background Technology

[0002] With the widespread adoption of mobile internet and smart devices, facial expression recognition solutions have emerged in large numbers, offering significant potential applications in fields such as healthcare, social media analytics, and personalized recommendations. However, current facial expression recognition methods often rely on raw images or feature data for identification, which carries the risk of personal privacy leaks. Homomorphic encryption, as a special encryption technique, allows computational operations, such as addition and multiplication, to be performed in ciphertext without decryption. This is extremely useful for performing forward and backward propagation operations of deep learning models within ciphertext. Facial expression recognition faces challenges in several aspects, including computational complexity, performance overhead, model training, and inference efficiency. Therefore, continuous algorithm improvement and the development of more efficient encryption schemes are necessary. Current facial emotion privacy data recognition methods suffer from several problems: 1. Application scenarios are typically limited to the same model context, with strict requirements on input and output; 2. Data protection is relatively weak, often leaking size relationships during distance calculations or comparisons; 3. The computational inference process is inefficient, with significant computational time and communication overhead. These problems need to be addressed through continuous algorithm improvement, enhanced data protection, and optimized computational efficiency. Summary of the Invention

[0003] The purpose of this invention is to address the problems of existing facial expression recognition schemes, which typically require the collection of user facial images or plaintext facial feature values. Furthermore, traditional facial expression recognition processes suffer from limitations in application scenarios, and existing privacy-preserving computational facial expression recognition methods suffer from weak security, low computational efficiency, and high communication volume. This invention proposes a data privacy-preserving facial emotion recognition method for mobile group perception. It employs homomorphic encryption technology, allowing feature vector matching in encrypted form, thereby improving computational efficiency and data privacy protection, reducing the risk of personal privacy leakage, and providing a higher degree of privacy protection. It also offers high application flexibility, adapting to the needs of various application scenarios, including healthcare, social media analysis, and personalized recommendations.

[0004] To achieve the above objectives, the technical solution provided by this invention is as follows: a facial emotion recognition method for data privacy protection in mobile group perception, comprising a facial feature vector provider, two cloud servers (CP and CSP) providing powerful computing capabilities and storing encrypted data, and multiple mobile devices for facial emotion feature recognition; it includes the following steps:

[0005] Step 1: The face feature vector provider generates a PaillierTD key pair (pk, sk), where pk is the public key and sk is the private key. The private key sk is split into two parts: key 1 (sk1) and key 2 (sk2). sk1 and sk2 satisfy the following equations: sk1 + sk2 ≡ 0 (mod sk) and sk1 + sk2 ≡ 1 (mod N). The Chinese Remainder Theorem is used to solve these equations, resulting in a key pair (pk, sk1, sk2). Neither the split private key sk1 nor sk2 can completely decrypt the data encrypted with the public key pk. The private key sk1 is sent to the CP and the private key sk2 is sent to the CSP. The public key pk is distributed to the query users performing face emotion recognition.

[0006] Step 2: Construct a KD-tree data structure from all the face feature vectors with emotion tags provided by the face feature vector provider. The constructed tree data structure has multiple layers, and each node stores a portion of the face feature vectors. A node containing face feature vectors and the emotion tag corresponding to each vector is called a search space. Remove data from each leaf node to ensure that each search space on the KD-tree has a similar number of vectors. Sort the data on the nodes by column, encrypt all data using public key pk, and send it to the CP storing the data.

[0007] Step 3: The mobile device participating in facial emotion recognition has the ability to extract facial feature vectors, and is equipped with a camera and lightweight model inference capabilities. It extracts features from the captured or selected image to obtain the feature vector v of the image to be identified by facial emotion recognition, and encrypts the feature vector v with the public key pk and records it as the query vector Ev, which is sent to CP as a parameter for the facial emotion recognition request.

[0008] Step 4: After receiving the request parameter Ev, CP recursively calculates the distance between Ev and the first vector stored in all search spaces at the current level on the KD-tree. It selects the search space with the smallest distance to Ev and repeats the above calculation and selection process until the current search space is indivisible. The search space S at this point is the target search space. Search space S contains multiple face feature vectors. A threshold is calculated by calculating the absolute value of each column, defined as a value val. val is a value in the corresponding column of the current search space S, and is not less than the smallest value val in the corresponding column of the current query request. This value can be used for ciphertext comparison calculations. The method is obtained by comparing the values ​​of the current column with each value of the current column and taking the difference to obtain the distance contribution value of the current column to the face feature vector. The values ​​of the corresponding face feature vectors in each column are summed in encrypted form to obtain the distance array dis between the query vector Ev and each face feature vector in the search space. The k smallest vectors in dis are extracted. If there are fewer than k, the first face feature vector in the search interval S is taken as the reference vector. The weighted average of all expression probabilities is calculated according to the distance of the k closest face feature vectors. The expression recognition result is output as ans.

[0009] Step 5: The CSP generates a random noise number r and sends it to the CP. The CP adds the noise r to the expression recognition result to obtain the result. The CP and CSP simultaneously decrypt the value ans+r. The CP returns the decryption result of ans+r to the querying user, and the CSP returns the noise r. The querying user removes the noise to obtain the final expression recognition result.

[0010] Furthermore, in step 1, the key generation process is as follows: First, two prime numbers p' and q' need to be found. Each prime number has a KeyLen bit in binary, which represents the security factor. The longer the KeyLen, the more difficult it is to crack. Calculate the hidden values ​​p and q, where p = 2p' + 1 and q = 2q' + 1. Then, calculate the modulus N = pq, and the greatest common multiple of p-1 and q-1 λ = lcm(p-1, q-1). Calculate the inverse of the greatest common multiple of p-1 and q-1 with respect to N, μ = λ -1 mod N, we get the ciphertext generator g-N+1, with public key pk = (g, N) and private key sk = λ.

[0011] Further, in step 1, the private key λ is divided into two parts sk1 and sk2, which satisfy the following equation requirements: sk1+sk2≡0(mod sk), sk1+sk2≡1(mod N). According to the operation property of congruence equations, multiplying the above two formulas obtains the characteristic equation δ=sk1+sk2=sk·μmod(sk·N). According to the Chinese Remainder Theorem, it is required that δ≡0(mod sk) and δ≡1(mod N) hold simultaneously. Then sk1 is set as a σ-bit random number R, that is, sk1=R, sk2=sk·μ+η·skN-sk1, where η is a non-negative integer.

[0012] Further, in step 4, the ciphertext comparison algorithm used in the binary search process is divided into three steps: In the first step, the CP generates a random number π, which takes a value of 0 or 1. If π=0, the CP calculates the difference D=r1-r2 between the two numbers to be compared r1 and r2; otherwise, it calculates D=r2-r1. The comparison difference D is partially decrypted by using sk1 to obtain the decryption result D1, and the CP sends (D, D1) to the CSP. In the second step, the CSP uses sk2 to partially decrypt D to obtain D2, and uses the partial decryption results D1 and D2 to perform final decryption to obtain the plaintext d of D, d=D1*D2mod N 2 and compares d with N / 2 in size. If d>N / 2, then d=1; otherwise, d=0, and returns d to the CP. In the third step, the comparison result, that is, the recognition result, is obtained by calculation.

[0013] Further, in step 4, the process of extracting the k smallest vectors from dis is as follows: First, initialize the left and right intervals as the minimum value low=Min and the maximum value high=Max of the value range of the current column. When low<high, compare the value at the (high+low) / 2-th position with val, and the comparison process adopts the ciphertext comparison algorithm. If the value is less than val, set low=(high+low) / 2; otherwise, set high=(high+low) / 2. The comparison process is repeated until low>high, and low at this time is the value d.

[0014] Further, the distance of the face feature vector adopts Manhattan distance, and the calculation formula is: Manhattan distance=∑x i -y i , wherein x i , y i respectively represent the coordinates of two points in the i-th dimension, and this formula represents the sum of the absolute values of the coordinate differences of the two points in each dimension, thereby obtaining the Manhattan distance between them.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0016] 1. It can dynamically adapt to various application scenarios for facial expression recognition without being limited by the input model; the group perception system represents a large network of people with interconnected relationships, where each entity can independently complete data extraction. It is usually equipped with cameras and lightweight model inference functions. Each entity participating in facial expression recognition has its own local feature extraction and computing capabilities. Through taking pictures or selecting images, these devices can extract facial features. These entities can be individuals distributed in different geographical locations, industries, and fields.

[0017] 2. By using binary search, the number of calculations in nonlinear ciphertext operations is effectively reduced, which can effectively improve the computational efficiency and reduce communication overhead.

[0018] 3. It can simultaneously protect the data security of model providers (data outsourcing users) and query users. Specifically, unlike traditional facial expression recognition systems, the method of this invention uses fully encrypted data transmission and computation, and achieves secure computation of sensitive facial feature data through PaillierTD's additive homomorphism.

[0019] In summary, the method of this invention employs homomorphic encryption technology, which allows feature vector matching in ciphertext, thereby improving computational efficiency and data privacy protection, reducing the risk of personal privacy leakage, and providing a higher degree of privacy protection. It has high application flexibility and can adapt to the needs of different application scenarios, including healthcare, social media analysis, and personalized recommendations, and is worthy of promotion. Attached Figure Description

[0020] Figure 1 This is a framework diagram of the method of the present invention.

[0021] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto. It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0023] like Figure 1 and Figure 2As shown, this embodiment provides a facial emotion recognition method for data privacy protection in mobile group perception. It extracts and encrypts features by querying the user's local computing power, performs matching in the cloud via data outsourcing, and then recognizes facial expression data based on the matching results. The method involves a facial feature vector provider, two cloud servers (CP and CSP) providing powerful computing capabilities and storing encrypted data, and multiple mobile devices for facial emotion feature recognition. To facilitate understanding of the implementation of the technical solution, specific data is used for demonstration in this embodiment.

[0024] Step 1: The face feature vector provider generates a PaillierTD key pair (pk, sk), selecting p = 2293860749, q = 3276141289, and calculating n = 7515011911015365461, where pk = (7515011911015365461, 7515011911015365462), sk = 7515011905 445363424; Split the private key sk into sk1 = 14899335598462580858082299688271082824 and sk2 = 21095095585612241498107814197332700952. Send part of the private key sk1 to the CP and part of the private key sk2 to the CSP. Distribute the public key pk to the users performing face emotion recognition queries.

[0025] Step 2: Construct a KD-tree data structure from the 20 5-dimensional facial feature vectors with emotion tags provided by the facial feature vector provider. The constructed tree data structure has multiple layers, and each node in each layer stores a portion of the facial feature vectors. A structure with two layers is constructed. The first layer's first space contains [1998,334,4036,2294,3767], [2136,3883,756,1295,952], and [2476,2733,1141,3664,2237], a total of 3 vectors. The second space in the first layer contains [3966,1208,1845,3130,3566] and [690,1987,461,3831,1984]. [1437,247,1949,1532,3688], a total of 3 vectors, the first space of the second layer has [852,1406,3149,2594,2847], [2096,2762,3965,1768,3439], [3805,3072,1413,148,1581], a total of 3 vectors, the second space of the second layer has [1904,1155,1348,1235,3030], [1242,1101,1617,1928,2201], [2757,2720,3944,1872,3027], a total of 3 vectors, the third space of the second layer has [3463,2205,2064,29... [50,3410],[44,2145,580,3056,1077],[1464,3697,1560,1,2646],[3802,3755,738,2418,1132], a total of 4 vectors. The fourth space of the second layer has [3802,3755,738,2418,1132],[657,621,2199,708,3363],[2971,1174,4018,232,2355],[3403,2995,588,3363,1193], a total of 4 vectors. After encryption with pk, the first space of the first layer has [339974409985896134800710]. 47348307052340,26307730514534567364364173622050776176,47588802159758751353659967938543883188,99112299594454330416929290581489 85620,31439778181623164065050177451776906600],[21329231794717611411302144308699211470,29931249172221075390898592794891023727,29853434283592569188735011092861935303,39281147066449498130437608237401290576,14712345831203357737443525088797129337],[38232592223410545588227478158596224874,48950074415990230224069921137163771124,47451508243642050320135517814375112251,8469447 440432649149204340570245885135,46932400671758660800720560750520273232], a total of 3 encryption vectors. The second space of the first layer contains [29196444627245361893911985602702304502,5840005733036900672184608856351817569,5127199545274021668299392764871014438,41467628401034207901697170152709873876,4 0486318617925269627510285118822776594],[37962615514164943325437968949035013380,25719962525599391416886356060051704795,8077062018996280264827631928719694336,30210742618149115688526399827753551007,11461674653002812348950309302537509627],[2906634 There are 3 encryption vectors in total. The first space of the second layer contains [354874866825008519575468267585621914861,35817243268593515038808225172600204528,6025753254732747093183699656963357181,41182506638680868544920195302326027479].53778186175085151431941723386013451616,31260853900216991135840822864183246169,12457418509185927236929112669770687068,2728443 0333289254898534637087017857820],[36460975255217064344881442923945674397,30693756084922358135936777972928572675,1210587762155 8050478135280073834333299,21516579064707782902464781568676946335,7969853144523183609671420188129756824],[2827990259716942811 5717296652717918686,5601349233685311003732138956245641927,17022862118242476838743199677487326237,3529864927134212074200007559 There are 3 encryption vectors: 6458077661,45170404454385643808463523393236820147. The second space of the second layer contains [28668320475224127845558680131672581989,1977345935516134]. 9147845731431359686440,44971808244541930844498081679141053323,23853003955126039756944702111716410958,583512702413532217568978 5570270339885],[42153487363809638105960645049445469418,29770863588273206603068548953723370988,4976391218073159750453858616542 4666106,9824716309770443415124717468330000553,6922180880225377981952569514343539248],[19111404491663034188634596545893622622,There are 3 encryption vectors in total. The third space of the second layer contains [52048960936854444760108834413767952910, 430518592782595172037835828]. 50258611491,6141663949328140061831425864798801278,45034996287102006178489467325612823622,23305158445915766579736999084307658463],[18677726932256248082832804413855512022,26129052163013559097936876777030267672,103038576366260306598450518049074 01094,11127942689069163742808480722016015107,14434241185841456591176886053731800592],[36499118198138809261116897454453174702,4362678798457801528171776325602939449,8865360859471536996593835759235439911,50008892246510758411356329388837381318,1 7193501451217781290524001767961338244], [56353231225672438666324495288532934019, 18419125046889055331694312040365520904, 811153063741044724809177627635789248, 35961949998126443748421873371725862480, 47633370739759501818966613761769848929], a total of 4 encryption vectors,The fourth space on the second floor contains [56353231225672438666324495288532934019,18419125046889055331694312040365520904,811153063741044724809177627635789248,35961949998126443748421873371725862480,476333707397595018189666137617698489]. 29],[11965192444880090116008442537742336687,32233852245437351871801073799816335109,28246907336157187992258874405495484684,26517061826173025819926072970177583430,18267866653371123845711801005201936377] ,[6492534643526080594457096046280440328,6513717279591472870603555141253869771,46160928008945259114274786556399092443,49457903444593990408013073413180932953,32290721438819222158560638444840201374],[486 [89615890456621711316855803327672007,10980672194965103975487203496443835727,21194139264959225051817099194926874429,13421438936594261944426550488904806655,53968029085884189780077219869511045613], a total of 4 encryption vectors.

[0026] Step 3: The mobile device participating in facial emotion recognition has the ability to extract facial feature vectors, is equipped with a camera and lightweight model inference capabilities, and thus extracts the feature vector v = [1337,274,1524,53,3477] from the captured image. This feature vector v is then encrypted using a public key pk and recorded as the query vector Ev = [15262213804548564443555363120963121780,19469567871]. [847435632271808462558486529,50690266461718383098888058026922099248,18331261191687359080683901443591303317,50247180069904391028544607794868400773] will be sent to CP as parameters for the facial emotion feature recognition request.

[0027] Step 4: After receiving the request parameter Ev, CP recursively calculates the first layer distance as [7050, 6413]. It then enters the second layer of sub-regions 1 and 2 within region 1, calculating the second layer distance as [3253, 7561]. The current search space, layer 2, is the target search space, containing multiple face feature vectors: [35487486682500851957546826758562180958, 53778186175085151431941723386013451616, 3126085390021699113584082286418324]. 6169,12457418509185927236929112669770687068,27284430333289254898534637087017857820],[36460975255217064344881442923945674397,30693756084922358135936777972928572675,12105877621558050478135280073834333299,2151657906470778290246478156 8676946335,7969853144523183609671420188129756824],[28279902597169428115717296652717918686,5601349233685311003732138956245641927,17022862118242476838743199677487326237,35298649271342120742000075596458077661,45170404454385643808463 [523393236820147] Calculate the threshold by the absolute value of each column, and get a value val that is a value of the corresponding column in the current search space S. The value of val is obtained by binary search [2096,1406,3149,148,3477]. The values ​​of the corresponding face feature vectors of each column are summed in ciphertext to get [6413,7441,7368]. The smallest vector in dis is extracted [852,1406,3149,2594,2847]. The weighted average of all expression probabilities is calculated, and the expression recognition result is output as ans, which is actually happy, with an index of 1.

[0028] Step 5: The CSP generates a random noise number r = 3 and sends it to the CP. The CP adds the noise r to the expression recognition result to obtain the result. The CP and CSP simultaneously decrypt the value ans+r to obtain 3. The CP returns the decryption result 4 of ans+r to the querying user. The CSP returns the noise r=3. The querying user removes the noise and obtains the expression recognition result 4-3=1, thus obtaining the recognition result of happy.

[0029] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A facial emotion recognition method for data privacy protection in mobile group perception, characterized in that, This includes a facial feature vector provider, two cloud servers (CP and CSP) that provide powerful computing capabilities and store encrypted data, and multiple mobile devices for facial emotion recognition; It includes the following steps: Step 1: The face feature vector provider generates a PaillierTD key pair (pk, sk), where pk is the public key and sk is the private key. The private key sk is split into two parts: key 1 (sk1) and key 2 (sk2). sk1 and sk2 satisfy the following equations: sk1 + sk2 ≡ 0 (mod sk) and sk1 + sk2 ≡ 1 (mod N). The Chinese Remainder Theorem is used to solve these equations, resulting in a key pair (pk, sk1, sk2). Neither the split private key sk1 nor sk2 can completely decrypt the data encrypted with the public key pk. The private key sk1 is sent to the CP and the private key sk2 is sent to the CSP. The public key pk is distributed to the query users performing face emotion recognition. Step 2: Construct a KD-tree data structure from all the face feature vectors with emotion tags provided by the face feature vector provider. The constructed tree data structure has multiple layers, and each node stores a portion of the face feature vectors. A node containing face feature vectors and the emotion tag corresponding to each vector is called a search space. Remove data from each leaf node to ensure that each search space on the KD-tree has a similar number of vectors. Sort the data on the nodes by column, encrypt all data using public key pk, and send it to the CP storing the data. Step 3: The mobile device participating in facial emotion recognition has the ability to extract facial feature vectors, and is equipped with a camera and lightweight model inference capabilities. It extracts features from the captured or selected image to obtain the feature vector v of the image to be identified by facial emotion recognition, and encrypts the feature vector v with the public key pk and records it as the query vector Ev, which is sent to CP as a parameter for the facial emotion recognition request. Step 4: After the CP receives the request parameter Ev, recursively calculate the distance between the first vector stored in all search spaces at the current level on the KD-tree and Ev, select the search space with the smallest distance to Ev, repeat the above calculation and selection process until the current search space can no longer be divided. At this time, the search space S where it is located is the target search space. Search space S contains a plurality of face feature vectors, and the threshold is calculated by calculating the absolute value of each column, which is defined as a value val that satisfies that val is a value in the corresponding column of the current search space S, and is not less than the smallest val in the corresponding column of the current query request. This value can be obtained by comparison through a binary search method relying on the ciphertext comparison algorithm. The difference is obtained by comparing the size of val with each value in the current column, to obtain the distance contribution value of the current column to the face feature vector. Ciphertext summation is performed on the values corresponding to the face feature vectors of each column, so as to obtain the distance array dis between the query vector Ev and each face feature vector in the search space. Extract the k smallest vectors from dis. If the number is less than k, take the first face feature vector in the search interval S as the reference vector, use the distances of the k closest face feature vectors as weights, calculate the weighted average of all expression possibilities, and record the output expression recognition result as ans; Step 5: The CSP generates a random noise number r and sends it to the CP. The CP adds the noise r to the expression recognition result to obtain the result. The CP and CSP simultaneously decrypt the value ans+r. The CP returns the decryption result of ans+r to the querying user, and the CSP returns the noise r. The querying user removes the noise to obtain the final expression recognition result.

2. The facial emotion recognition method for data privacy protection in mobile group perception according to claim 1, characterized in that, In step 1, the key generation process is as follows: First, two prime numbers p' and q' need to be found. Each prime number has a KeyLen bit in binary, which represents the security factor. The longer the KeyLen, the more difficult it is to crack. Calculate the hidden values ​​p and q, where p = 2p' + 1 and q = 2q' + 1. Then, calculate the modulus N = pq, and the greatest common multiple of p-1 and q-1 λ = lcm(p-1, q-1). Calculate the inverse of the greatest common multiple of p-1 and q-1 with respect to N, μ = λ -1 Mod N, we get the ciphertext generator g = N + 1, the public key pk = (g, N), and the private key sk = λ.

3. The facial emotion recognition method for data privacy protection in mobile group perception according to claim 2, characterized in that, In step 1, the private key λ is divided into two parts sk1 and sk2, which satisfy the following equation requirements: sk1+sk2≡0(mod sk), sk1+sk2≡1(mod N). According to the operation property of congruence equation, multiply the above two equations to obtain the characteristic equation δ=sk1+sk2=sk·μmod(sk·N). According to the Chinese Remainder Theorem, δ≡0(mod sk) and δ≡1(mod N) are required to hold simultaneously. Then set sk1 to a σ-bit random number R, that is, sk1=R, sk2=sk·μ+η·skN-sk1, where η is a non-negative integer.

4. The facial emotion recognition method for data privacy protection in mobile group perception according to claim 3, characterized in that, In step 4, the ciphertext comparison algorithm used in the binary search process consists of three steps: First, CP generates a random number π, which can be either 0 or 1. If π = 0, CP calculates the difference D = r1 - r2 between the numbers r1 and r2 being compared; otherwise, it calculates D = r2 - r1. The difference D is partially decrypted using sk1 to obtain the decrypted result D1, and CP sends (D, D1) to CSP. Second, CSP partially decrypts D using sk2 to obtain D2. The partial decryption results D1 and D2 are then used to perform the final decryption to obtain the plaintext d = D1 * D2 mod N. 2 The process involves comparing d with N / 2; if d > N / 2, then d = 1; otherwise, d = 0, and the result is returned to CP. The third step is to calculate the comparison result, which is the recognition result.

5. A facial emotion recognition method for data privacy protection in mobile group perception according to claim 4, characterized in that, In step 4, the process of extracting the k smallest vectors from dis is as follows: first, initialize the left and right intervals to the minimum value low=Min and the maximum value high=Max of the value range of the current column. When low<high, compare the size of the value at the (high+low) / 2-th position with val, and the comparison process adopts a ciphertext comparison algorithm. If it is less than val, set low=(high+low) / 2, otherwise set high=(high+low) / 2, repeat the comparison process until low>high, and at this time low is the value d.

6. A facial emotion recognition method for data privacy protection in mobile group perception according to claim 5, characterized in that, The distance between the facial feature vectors is calculated using Manhattan distance, which is: Manhattan distance = ∑x i -y i , where x i ,y i Let represent the coordinates of two points in the i-th dimension. This formula represents the sum of the absolute values ​​of the differences between the coordinates of the two points in each dimension, thus yielding the Manhattan distance between them.

Images