Scientific research recommendation system can provide the valuable reference for researchers to choose topics and determine research direction. However, traditional scientific research recommendation obtains the model by training the behaviour dataset of researchers stored in the centre, which may lead to the disclosure of researchers' sensitive information. In this paper, we propose a scientific research recommendation system based on privacy-preserving training dataset. Specifically, we use the federated learning mechanism and threshold homomorphic encryption technology to make the scientific research recommendation model available without uploading the raw dataset, which can protect the privacy of the researchers' behaviour dataset. Additionally, we also use a method to process the dataset of low-quality researchers to improve the accuracy of recommendation model. Through analysis, not only the researchers' privacy can be protected, but also the recommendation model accuracy can be optimized. The experimental results show that the proposed scheme can satisfy the functional requirements of scientific research recommendation system.
With the massive growth of scientific research information data, scientific research
recommendation system will become a right-hand man for researchers to choose their own scientific research
interests [
The traditional recommendation system centralizes the dataset to a central server for training. In
this way, the centralized storage of researcher’s behaviour dataset on the server may lead to the risk of
disclosure of private information. The reason is that the data information can be easily obtained by
malicious third parties or untrusted cloud servers. In order to overcome these problems, many scholars
have proposed to use federated learning mechanism to train the model [
To combat that, we propose a scientific research recommendation system based on privacypreserving training dataset. First of all, we test the local gradient and rule out the unreliable ones. Of course, we will ensure that a certain number of dataset are used to train the recommendation model. The contributions of this article can be summarized as follows: ⚫ First, we propose a scientific research recommendation system based on privacy-preserving training dataset. This scheme uses federal learning mechanism and threshold homomorphic encryption technology to protect the researchers’ privacy. ⚫ Second, the proposed scheme can mitigate the negative impact of low-quality data caused by unreliable researchers. ⚫ Finally, we also conduct a large number of experiments to verify that the proposed scheme has better performance in terms of security and efficiency.
The rest of this article is organized as follows. In Section 2, we describe the relevant primitives that this scheme needs to use. We introduce the system model and specific scheme in Section 3 and Section 4, respectively. We present the security and performance analysis in Section 5. Finally, we summarize the proposed scheme.
In the proposed scheme, we use the (t, n) threshold Paillier cryptosystem [
The homomorphic property of the above algorithm is as follows: c = Epk (mi + m j ) = g (mi +mj ) (rirj )ns mod ns+1 = Epk (mi ) Epk (m j ) (1)
Traditional machine learning is to train dataset together, so that it is possible to leak the raw data to
adversaries. To combat this privacy issue, Google first proposed a framework for federated learning in
2016, which allows distributed users to train locally without exposing their raw data. Federated
learning is technology that uses distributed optimization methods to protect data privacy in
multiparty cooperation [
As shown in Figure 1, the system model of the proposed scheme includes three entities, namely a trusted third party (TTP), a central server (CS) and researchers who provide dataset. Each researcher computes the local gradient by training his/her behaviour dataset locally, and then uploads the gradient to CS. After that, CS aggregates all uploaded gradients to train a global research recommendation model. At the same time, the global model is fed back to each local researcher, and they train the new gradient according to the global. The above iteration does not end until the accuracy of the global model meets certain requirements. The entities in the system model are described as follows:
We use cosine similarity to compare the correlation between local gradient and ideal gradient. The initial ideal gradient is a preset initial value.
In the proposed scheme, the threat model comes from external adversaries and internal adversaries. Central server can become internal adversary if corrupted by adversaries. It is possible to use its convenience to obtain the researchers’ behaviour information stored. Based on the given threat model, the requirement of the proposed scheme is to protect the privacy of gradient information provided by researchers, that is, the sensitive information will not be disclosed in the process of gradient transmission and storage. In addition, to improve the accuracy of recommendation system, unreliable participants should be screened before uploading the gradients.
In this section, we introduce the proposed scheme. First, our scheme considers the problem of unreliable researchers, that is, local gradient generated by ith iteration must be compared with the ideal gradient of this round in order to improve the accuracy of the recommendation model. Then, the security of gradient in transmission and storage procedures is also considered. The specific scheme includes the following four phases: system initialization, processing of low-quality dataset researcher, local gradient encryption and generation of recommendation model.
TTA is responsible for initializing the system. Given a security parameter , TTA generates a public key pk for all entities and assigns a set of private keys {sk1, sk2 , , ski , , skI } to each researcher Ri . G* = {G0*, G1*,
* , Gi ,
, GI*−1} is a global ideal gradient, which is generated by pretraining the scientific research recommendation model. Here, Gi* denotes the ideal gradient of the ith iteration.
Researchers with low-quality dataset should be screened before uploading local gradients. Otherwise, they will affect the accuracy of the scientific research recommendation system. Suppose G j = {G1j , G j , 2 , Gij , , G j} is the jth researcher’s local gradient, and i represents the number of
I iterations. Given the ideal gradient of ith iteration Gi* , each participating researcher compares its own gradient with it, as follows: sim(Gi*−1, Gij ) =
Gi*−1 Gij Gi−1 Gij * (2) where, sim() is the cosine similarity algorithm. According to equation (2), the higher the value of sim() , the higher the reliability of local gradient. When the result of sim() is less than a certain value, it indicates that the researcher with the gradient is an unreliable participant. In the proposed scheme, it is assumed that I pariticipating researchers are needed to train the global recommendation system. Unreliable researchers are screened and new participants are reselected to ensure that a certain number of researchers come to train global model. The specific process is illustrated in Algorithm 1.
Each reliable participant j encrypts his/her gradient as c j = Encpk (G j ) with a public key pk and then uploads it to the central server. After receiving all encrypted local gradients, the central server aggregates all encrypted gradients as follows: c = c1c2...cn = Encpk (G1)Encpk (G2 )...Encpk (Gn )
= Encpk (G1 + G2 + ... + Gn ) Algorithm 1: Processing of Low-quality Data Researcher Input: Global ideal gradient G* = {G0*, G1*,
* , Gi , , Gij , , GI*−1} , , G j}
I local gradient
G j = {G1j , G j ,
2 threshold TH Output: reliable gradient
* 1: Initialize global ideal gradient G ;
* 2: In jth iteration, given G ; 3: for i to I do 4: if Eqn. ( 2) TH then 5: Return Gij ; 6: end if 7: end for
The recommendation model is derived from the aggregate values of all upload gradients. Therefore,
the central server needs t participants to use their private keys sk j to calculate the secret share
c j = c2sk j mod ns+1. Then, using t shares of c j , the plaintext of aggregated gradients can be restored
by Lagrange interpolation algorithm [
In this section, we will discuss the security and performance of the proposed scheme. Additionally, the experiments are based on MNIST database and carried out on an operating system with intel (R) Core (TM) i7-9750H and 8G RAM.
The security of the proposed scheme focuses on how to protect the privacy of researchers who provide training dataset. Specifically, the participating researcher’s gradient is protected.
Proof: In the proposed scheme, the researchers involved in the training do not send their raw data to the central server, but only trained locally. Therefore, adversary will not be able to obtain the raw information of the researchers. In addition, each trained gradient is encrypted by threshold Paillier cryptosystem as c j = Encpk (G j ) . From Section 4, the decryption key needs to be recovered by at least t participating researchers, so it is very difficult for adversary and central server to obtain the decryption key. Therefore, our scheme can protect researcher’s dataset.
Here, we mainly discuss the computation cost of encryption and decryption, and the efficiency of low-quality user verification in the proposed scheme.
In local gradient encryption phase, each researcher encrypts his/her gradient as c j = Encpk (G j ) . And the recovery of global gradient is recovered in the model generation phase.
Figure 2 shows the computation cost of gradient encryption and decryption with the number of iterations. Next, we discuss the accuracy of the proposed scheme after considering the processing of low-quality dataset researchers. In order to better describe the experiments, we compare the proposed scheme with normal federated learning mechanism (NFM) scheme and filtered but not reselected federal learning mechanism (FBNRF) scheme in terms of accuracy. NFM refers to a federated learning scheme that does not deal with low-quality dataset researchers, and FBNRF denotes that lowquality dataset researchers have been screened but have not been reselected.
As shown in Figure 3, since the ideal global gradient is not obtained, the accuracy in the three scenarios in the initial iteration is about 10%. However, after three iterations, we can see that the accuracy of FBNRF and our scheme is better than that of NFM. In the 9th iterations, the accuracy of the three scenarios is 72%, 84%, 95%, respectively. And after completing the number of iterations, the accuracy of our scheme is higher than that of the other two schemes. 6.
In this article, we propose a scientific research recommendation system based on privacypreserving training dataset, we test the local gradient and rule out the unreliable ones and use the federated learning mechanism and threshold homomorphic encryption technology to make the scientific research recommendation model available without uploading the raw dataset. Security and performance analysis shows that the proposed scheme can meet the security and efficiency of dataset of scientific research recommendation system. In the future, we will study the privacy protection of query users and model parameters in the scientific research recommendation.