3.2.1. Data pre-processing.

This study uses the RAPPOR method in LDP to pre-process the raw data to obtain the ε-localized differential privacy-protected data D to obfuscate student identity information. Importantly, LDP is applied only to metadata and behavioral data (e.g., forum participation frequency, time spent on modules) to preserve aggregate utility for analytics, while academic records (e.g., grades, quiz scores) remain unperturbed and are secured via ABSE. This separation ensures that instructors can retrieve precise student grades when authorized, while LDP safeguards metadata against identity linkage. RAPPOR's two-stage randomization (PRR and IRR) introduces controlled noise into metadata, preventing the disclosure of personally identifiable information (PII) without compromising the integrity of academic records.

RAPPOR is a pivotal component of the data pre-processing phase, as it anonymizes student responses by adding controlled noise, preserving overall data patterns while protecting individual privacy.

A total of 4 steps are included: bloom filtering (BF), permanent random response (PRR), instantaneous random response (IRR), and aggregation. The data pre-processing is shown in Fig 4.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Data pre-processing.

https://doi.org/10.1371/journal.pone.0347786.g004

1. (1). BF. We use the BF technique to represent the actual data m as a vector of length b. This step is the starting point for data preprocessing and lays the foundation for subsequent random response mechanisms. At the initial state, for an array of string length z bits, its h mapping function is set to b 1's in the bitmap. The value of real data m is represented as a vector of length b by BF technique, and the mapping relationship between Bloom strings and strings is recorded.
2. (2). PRR. In the PRR step, we perturb each bit of the vector W to generate a new vector W.’ This process introduces a degree of randomness to protect the privacy of individual data. For the real data m, a new vector W’ is generated by perturbing each bit of the vector W. f ∈ [0,1] denotes the probability of perturbation, such that it satisfies equation:

(1)

1. (3). IRR. The IRR step further perturbs each bit of W’ to produce the final binary string result set D, which represents the uploaded data. A second perturbation of each bit x of the vector W’ generates a z-bit binary string result set D, representing the probability of being set to 1 when it takes the value 1 and 0, respectively.

(2)

Because the instantaneous random response is a second perturbation process, the 1 in the uploaded data can be transformed from 0 in the original data or from 1 in the original data. If it is transformed from 1 in the original data, the probability of such a transformation is denoted as *. This transformation can be composed of 2 kinds of changes: from 1 to 0 and then become 1; or from 1 to 1 and then become 1. Then,

(3)

If it is transformed from 0 in the original data, the probability of this transformation is denoted as q*. This transformation can also consist of 2 changes: from 0 to 0 and then to 1; or from 0 to 1 and then to 1. Then,

(4)

1. (4). Aggregation. The sensitive data cipher text D generated after the instantaneous random response operation is uploaded to the CS.

3.2.2. System establishment.

In the ABE algorithm, each user's attribute set is different. When the resource owner uploads shared resource data, in order to achieve permission control for different users, an access policy is formulated based on the user's identity attribute set. Only users who meet the access policy can access the data. Commonly used access strategies are Boolean expressions, linear secret sharing matrices and access trees. The access policy chosen in this study is linear secret sharing matrix. In addition to this, this paper combines the SE technique and ABE technique so that the data privacy protection and access control problems are solved.

: The algorithm is executed by AAC with input parameter, which produces the bilinear mapping . Where, A and are cyclic groups of order prime , and a is the generator of A. To initialize the bilinear mapping, we first select a large prime and choose two generators. Then, we use these generators to define the bilinear mapping e.

The generation of random numbers during system initialization is a key step to ensure the security of keys and other parameters. We use a high-quality random number generator to generate the necessary random numbers. These random numbers are used to generate the master key, user keys, and other necessary system parameters. The specific steps are as follows: 1) Use a secure random number generator to generate random numbers . 2) Use these random numbers to compute the system's public key parameters and master key.

(5)

(6)

: Key generation. The algorithm is executed by the AAC with inputs of the master keys MSK, PP, and the set of user attributes S. For each attribute , the algorithm randomly selects the parameter , and computes . Generate the user decryption key SK.

(7)

3.2.3. Encryption and uploading.

In the data encryption and upload stage, teachers use public parameters PP, plaintext w and access policies to implement data encryption algorithms. Here are the detailed steps and explanations of the relevant mathematical equations:

: Data encryption. The algorithm is executed by the teacher with input public parameters PP, plaintext w and access policy . Where, G denotes a l × t matrix, denotes the mapping function that maps each row of G to the corresponding attribute, and denotes the shared sub-secret for row . The algorithm randomly selects vectors . For each row of G, compute the shared sub-secret , randomly selecting parameters . Compute the data cipher as follows.

(8)

For the set of keywords , the data containing the keyword is generated as an index . The parameter is chosen randomly and the keyword index is computed as

(9)

The teacher uploads the data cipher text to the CS, which then returns its storage address to the teacher; and broadcasts the storage of the index into a new block of the blockchain.

Algorithm 1: Data encryption and uploading process

1. (1). Input: The teacher initiates the algorithm with the public parameters PP, plaintext w, and the access policy.
2. (2). Matrix G creation: Generate a matrix G that represents the attributes associated with the data.
3. (3). Random vector selection: Select random vectors for the encryption process.
4. (4). Shared Sub-secret calculation: Compute the shared sub-secrets for each attribute mapped to a row in matrix G.
5. (5). Data cipher computation: Calculate the data cipher using the shared sub-secrets and the plaintext w.
6. (6). Index creation for keywords: For each keyword in the data, create an index using a hash function and a randomly chosen parameter.
7. (7). Data upload to CS: Upload the encrypted data to the CS.
8. (8). Block chain index broadcast: Broadcast the keyword index to the block chain for data retrieval purposes.

3.2.4. Data query and matching.

: Trapdoor generation. The algorithm is executed by the students, which inputs the public parameter (PP), the individual private key SK and the keyword m’ for the search. The algorithm randomly selects the parameter . The trapdoor is generated and sent to the blockchain.

(10)

Search: Search for matching. The algorithm is executed by a blockchain node with inputs of public parameter PP, keyword index and user trapdoor . If the search keyword is the same as the keyword contained in the index, Equation (11) is satisfied, indicating that the search match is successful and the blockchain returns the address to the teacher.

(11)

3.2.5. Download decryption.

Decrpt : ciphertext decryption. The algorithm is executed by students with the input public parameter PP, the ciphertext under the access policy , and the decryption key under the set of attributes S, as follows, respectively.

(12)

(13)

If the set of attributes S satisfies the access policy , then compute the value that satisfies

(14)

The decryption algorithm is as follows.

(15)

Finally, the algorithm computes to obtain the plaintext m.

The threat models that this proposal can counter include the capabilities and intentions of potential adversaries, such as passive eavesdroppers and active attackers capable of launching various attacks, such as data injection, denial, and privacy breaches.

The scheme is designed to achieve several critical security properties, as follows:

1. (1). Confidentiality: Utilizes ABE to ensure that only authorized entities (teachers and students) with specific attributes can access encrypted student data stored in the CS.
2. (2). Integrity: Maintains data integrity through cryptographic techniques such as digital signatures and hashing, ensuring that data stored on the blockchain cannot be tampered with or altered without detection.
3. (3). Privacy preservation: Implements LDP using methods like RAPPOR to obfuscate individual student responses before aggregation, thereby protecting the privacy of student identities and sensitive data.
4. (4). Access control: Enforces fine-grained access control through ABE, where access policies based on attributes ensure that only authorized users can decrypt specific data based on their roles and permissions.
5. (5). Resilience to attacks: The use of blockchain technology enhances resilience against attacks by providing decentralized and tamper-resistant storage for data indexes and transaction records, reducing the risk of single points of failure.

These properties are formally grounded in the Decisional Bilinear Diffie-Hellman Exponent (BDHE) assumption for ABSE confidentiality and epsilon-differential privacy for LDP, providing provable security against chosen-plaintext attacks (CPA) and inference threats (e.g., membership inference attacks), respectively.

4.2.1. Privacy budget consumption.

Our choice to use MNIST and CIFAR-10 datasets allows us to benchmark the privacy budget consumption of our proposed method against existing algorithms in a controlled setting. Specifically, we aim to measure how effectively our method maintains privacy (as quantified by the privacy budget) while still allowing for accurate data analysis. This measurement is crucial for understanding the practical applicability of our method in VCLEs, where privacy must be balanced with the need for educational analytics.

To gauge the efficacy of LDP in curtailing privacy budget expenditure within this algorithm, the algorithms outlined in literatures [21] and [22] are employed for juxtaposition. The privacy budget utilized by the three algorithms to achieve the designated precision is documented for the MNIST and CIFAR10 datasets, depicted in Figs 5 and 6, respectively. Where ε[21], ε[22], and εP represent the privacy budget consumed by the literature [21] algorithm and literature [22] algorithm, and the proposed algorithm, respectively.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Privacy budget consumed by different algorithms on MNIST dataset.

https://doi.org/10.1371/journal.pone.0347786.g005

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Privacy budget consumed by different algorithms on CIFAR10 dataset.

https://doi.org/10.1371/journal.pone.0347786.g006

As can be seen from Fig 6, when the accuracy is 95%, the proposed algorithm reduces the privacy budget by 1.7% and 1.9% compared to the methods in [21] and [22], respectively. When the accuracy is 96%, the proposed algorithm reduces the privacy budget by 1.6% and 2.0% compared to literature [21] and literature [22], respectively. When the accuracy is 98%, the proposed algorithm reduces the privacy budget by 2.8% and 2.4% over literature [21] and literature [22], respectively. From Fig 6, it can be seen that the privacy budget of the proposed algorithm is reduced than that of the other two compared algorithms at the same accuracy. The experimental results show that the proposed algorithm consistently outperforms literature [21] and [22] in terms of privacy budget consumption at the same accuracy level. This is primarily due to the effective use of LDP, which ensures that privacy is preserved without excessive budget usage. As the results indicate, the privacy budget required to achieve high accuracy is lower in our algorithm compared to the other two methods. This is important because a smaller privacy budget signifies a higher level of privacy protection, which is a critical factor in applications such as VCLEs, where maintaining user privacy is paramount.

To validate the proposed method's performance on diverse educational data, we evaluated its privacy budget consumption on the student behavior dataset and learning progress dataset (see Table 1). For the student behavior dataset, we set ε = 3 to balance privacy and utility, given the high-dimensional and unstructured nature of the data. The proposed algorithm achieved a privacy budget reduction of 2.1% and 2.4% compared to literature [21] and [22], respectively, at an accuracy of 94%. For the Learning Progress Dataset, with ε = 3.5 to accommodate the time-series data, the proposed algorithm reduced the privacy budget by 1.8% and 2.4% compared to literature [21] and [22], respectively, at an accuracy of 92%. These results, presented in Table 1, demonstrate that the proposed method maintains strong privacy protection across complex educational datasets while preserving data utility for analytics, such as identifying engagement patterns or tracking academic progress.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Privacy budget consumption comparison for educational datasets.

https://doi.org/10.1371/journal.pone.0347786.t001

4.2.2. Accuracy of the algorithms.

The accuracy of our proposed algorithm on the MNIST and CIFAR-10 datasets is a key performance indicator that reflects its potential effectiveness in a VCLE setting. Because they provide a standardized benchmark for evaluating the trade-off between privacy preservation and data utility. In the context of VCLE, accuracy can be seen as a measure of how well educational data can still be used for its intended purpose (such as adaptive learning and educational analytics) after privacy enhancements have been applied. To assess the influence of incorporating LDP into the blockchain on the algorithm's precision, the proposed algorithm is compared with the other two comparative algorithms in terms of accuracy rate, given the same privacy budget. The results are shown in Fig 7 and Fig 8.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Comparison of accuracy of three algorithms on MNIST dataset.

https://doi.org/10.1371/journal.pone.0347786.g007

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Comparison of accuracy of three algorithms on CIFAR10 dataset.

https://doi.org/10.1371/journal.pone.0347786.g008

According to Fig 7, it can be seen that on the MNIST dataset, due to the limitation of privacy budget ε = 2, the proposed algorithm is trained to stop at 35 rounds with an accuracy of 99%; literature [21] algorithm stops when it is trained to 25 rounds with an accuracy of 92.5%; literature [22] algorithm is trained to stop at 20 rounds with an accuracy of 91%.

As shown in Fig 8, on the CIFAR10 dataset, due to the limitation of the privacy budget ε = 4, the proposed algorithm stops when it is trained to 60 rounds with 73% accuracy; literature [21] algorithm stops when it is trained to 50 rounds with an accuracy of 47%; literature [22] algorithm stops when it is trained to 40 rounds, and the accuracy rate is 42%. It can be seen that by introducing LDP technology in the blockchain, the proposed algorithm is able to train more rounds under the same privacy budget and achieve a higher accuracy rate. Therefore, the proposed algorithm is suitable for virtual learning environments that require high accuracy and privacy protection.

A comprehensive comparison of model performance under different privacy budgets (ε) is summarized in Table 2. On the student behavior dataset, with ε = 3, the proposed algorithm achieved an accuracy of 94% after 40 training rounds, compared to 88% for literature [21] (30 rounds) and 86% for literature [22] (28 rounds). For the learning progress dataset, with ε = 3.5, the proposed algorithm reached 92% accuracy after 45 rounds, outperforming literature [21] (85%, 35 rounds) and literature [22] (83%, 32 rounds). These results, presented in Table 2, confirm that the proposed method maintains high accuracy across diverse educational data types, supporting its applicability in real-world VCLEs for tasks such as personalized learning recommendations and progress monitoring.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Accuracy and training rounds comparison for educational datasets.

https://doi.org/10.1371/journal.pone.0347786.t002

4.2.3. Transaction per second.

System transaction per second is an important factor in evaluating the performance of a blockchain, in this case the number of keyword transactions processed per second. The TPS metric serves as a performance indicator of our blockchain-based system's capability to process keyword transactions, analogous to the querying and retrieval of educational resources within a VCLE. A higher TPS reflects the system's efficiency in handling large volumes of data transactions while preserving privacy, an essential requirement for sustaining the responsiveness and usability of educational platforms. To enhance TPS, we optimized blockchain operations for keyword transactions by minimizing latency and maximizing throughput. These optimizations include implementing efficient indexing mechanisms and enabling parallel processing of transactions, both of which are critical for achieving elevated TPS in our system. The experimental test is conducted in the same experimental environment, the number of nodes in the network is 20, and each block contains 50 transactions. The TPS of the three methods were tested separately and the results are shown in Fig 9.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. TPS comparison of different algorithms.

https://doi.org/10.1371/journal.pone.0347786.g009

From Fig 9, it can be seen that the TPS of our method is higher than the other two methods. The reported 25 TPS reflects a baseline measurement from our experimental setup, which prioritized validating the privacy-preserving features of the framework over maximizing throughput. For large-scale educational platforms requiring higher TPS, scalability solutions such as sharding, off-chain processing, or layer-2 protocols (e.g., state channels) can be integrated to significantly boost performance. Given that this study focuses primarily on privacy and data security, the current TPS is sufficient for practical VCLE use cases, with future work planned to explore additional optimizations for broader scalability.

4.2.4. Scalability analysis.

Scalability challenges in our blockchain-based framework for VCLEs arise primarily from network growth and transaction volume, such as congestion in PoA consensus and block capacity limits, which can degrade performance in large-scale settings (e.g., > 10,000 students). In our setup with 20 nodes, baseline TPS reached 25, but simulations scaling to 100 nodes revealed a 15−20% TPS drop due to increased synchronization overhead, aligning with Ethereum's observed limits of ~15 TPS under similar loads. This highlights potential bottlenecks like node communication delays, critical for high-interaction VCLEs. To address these within the framework, sharding—partitioning the chain into parallel shards—can distribute load and boost throughput by 10-100x, while Layer-2 rollups offload computations to reduce main-chain congestion without security trade-offs. Our PoA choice mitigates some issues in permissioned educational networks by minimizing energy-intensive puzzles, but hybrid integrations (e.g., sharding with PoA) would further enhance resilience for dynamic data sharing.

4.2.5. Latency and resource consumption analysis.

The integration of ABSE and LDP introduces notable latency and computational overhead due to their cryptographic operations. For LDP via RAPPOR, the two-stage randomization (PRR and IRR) adds noise injection overhead, increasing processing time by approximately 20–30% per data preprocessing step, as it involves multiple hash computations and perturbations. ABSE contributes further through attribute matrix calculations and trapdoor generation, with encryption/decryption latencies averaging 150–300ms per transaction on our i5-6300HQ setup, scaling with attribute set size (e.g., 10 attributes). Resource consumption includes higher CPU utilization (up to 70% during encryption) and memory (200–500MB for large datasets), potentially impacting real-time VCLE interactions like data queries. Quantitatively, experiments on MNIST revealed an average end-to-end latency of 2.5 seconds for ABSE-encrypted searches, compared to 1.2 seconds without encryption, while LDP preprocessing increased computational cycles by 40% relative to baseline methods. Mitigations include batch processing for LDP and hardware acceleration (e.g., GPU for matrix ops), reducing overhead by 15–25% in simulations. These trade-offs are acceptable for privacy-critical educational data, where security outweighs minor delays.

4.2.6. Runtime.

In order to further explore the impact of the combination of blockchain and LDP on the running efficiency of the algorithm, the proposed algorithm is compared with the original LDP algorithm in terms of running time, and the results are shown in Table 3. Among them, the training of the proposed algorithm cuts off when the privacy budget is consumed, and the training of the original LDP algorithm cuts off when the algorithm converges.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Running time comparison between the proposed algorithm and the original LDP algorithm.

https://doi.org/10.1371/journal.pone.0347786.t003

From Table 3, when the model converges, for the MNIST dataset, the running time of the proposed algorithm is about 3.5 times that of the original LDP algorithm (1950 s versus 1075 s). For the CIFAR10 dataset, the running time of the proposed algorithm is about 2.4 times that of the original LDP algorithm, which is 1950 s and 812 s. It can be seen that the consensus and validation mechanism in the blockchain increases the running time of the algorithm, but at the same time provides more security and stability for the virtual learning environment. The running time of the algorithm, but at the same time provides higher security and stability for the virtual learning environment. The running time of the algorithm is increased, but it simultaneously provides enhanced security and stability for the virtual learning environment.

To evaluate the performance indicators between the proposed method and those in literatures [21] and [22], standard statistical tests were used to evaluate each indicator to quantify the differences and determine their statistical significance (see Table 4).

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Statistical analysis results – privacy budget consumption.

https://doi.org/10.1371/journal.pone.0347786.t004

Table 4 presents the results of a t-test comparing the privacy budget consumption of the proposed method against the two existing methods from literature [21] and [22]. A P-value of less than 0.05 indicates the statistical significance of the difference in privacy budget consumption between the proposed method and the methods in literatures [21] and [22].

4.2.7. Evaluation of practical application effect.

In this study, an experimental platform was developed for a virtual collaborative learning environment. The platform was developed to provide a controlled environment where the proposed privacy protection methods could be tested and evaluated. It aimed to replicate the dynamics of virtual learning, offering a space for students to engage in collaborative activities and interactions. The study involved two distinct groups: the experimental group and the control group. The experimental group utilized the innovative method proposed in this study, which leverages blockchain technology for data uploading and management. This integration includes node verification and a privacy protection module, enhancing the platform's security posture. The control group followed the conventional approach of uploading data directly to a centralized server without the use of blockchain. This method served as a comparative baseline, highlighting the differences in privacy protection and security.

Data collected encompassed students’ private information, including personal details (A), academic performance metrics (B), health-related records (C), behavioral patterns (D), and communication data (E).The experimental group and control group were selected separately, and the experimental group used the proposed method to upload to the blockchain. The control group uses the traditional method and uploads directly to the server. Traditional method, i.e., a traditional approach that does not use blockchain technology and uploads data directly to a centralized server. This method serves as a baseline for comparison in this study. Fig 10 shows the difference in the probability of attackers identifying a student through a shared platform.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Success probability of attackers recognizing students’ real identities.

https://doi.org/10.1371/journal.pone.0347786.g010

The experimental results, shown in Fig 10, demonstrate that the proposed method significantly reduces the probability of attackers successfully identifying students’ real identities, particularly in sensitive data types such as personal details and communication data. This practical application analysis underscores the effectiveness of our approach in real-world scenarios, where it is crucial to protect student privacy while maintaining system performance. The integration of blockchain with node verification and privacy protection modules ensures that sensitive student data is safeguarded against unauthorized access and data breaches.

4.2.8. User and Instructor Feedback.

To assess usability, we conducted a pilot survey with 50 simulated users (30 students, 20 instructors) on our experimental VCLE platform. Feedback was collected via a 5-point Likert scale questionnaire focusing on privacy perception, ease of use, and overall satisfaction. Results showed high privacy satisfaction (average 4.6/5), with users noting reduced anxiety over data breaches compared to traditional systems. Instructors appreciated fine-grained access control (4.4/5), though some reported minor usability issues like initial setup complexity (3.8/5). Qualitative comments highlighted improved engagement due to perceived security, aligning with similar blockchain-based educational systems. This feedback validates the model's practical value, with suggestions for UI enhancements in future iterations.

4.2.9. Evaluation of anti-attack capabilities.

To comprehensively verify the privacy protection effectiveness of the proposed method, we conducted experiments simulating three common attack scenarios—data interception, unauthorized access, and collusion attacks—to assess the system's resilience against privacy breaches in virtual collaborative learning environments (VCLEs). These tests directly evaluate the anti-attack capabilities of the proposed framework, the results are summarized in Table 5. Simulations involved 100–200 runs per scenario (Monte Carlo-style, with random seed initialization for reproducibility), with “success” defined as adversary recovering >50% of plaintext (data interception), accessing unauthorized records (unauthorized access), or reconstructing identities from colluded data (>30% accuracy). Adversary model: semi-honest with partial access to ciphertext/network traffic, as per standard threat models in privacy literature.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Success probability of attack scenarios.

https://doi.org/10.1371/journal.pone.0347786.t005

Data interception attack: The proposed method, utilizing ABSE, prevents intercepted data from being deciphered without the necessary decryption keys, which are tied to user-specific attributes. In 100 simulated interception attempts on the MNIST dataset, the adversary's success rate in recovering plaintext was 0%, compared to 12% for literature [21] and 15% for literature [22], which rely on weaker encryption schemes. The resilience is largely attributed to ABSE's fine-grained access control and the blockchain's tamper-proof indexing, which effectively prevent key leakage.

Unauthorized access attack: In 200 simulated unauthorized access attempts on the CIFAR-10 dataset, the proposed method recorded a 0% success rate for unauthorized access, compared to 8% for literature [21] and 10% for literature [22], which use less stringent access controls. This result highlights the proposed method's effectiveness in maintaining data confidentiality.

Collusion attack: In 50 simulated collusion attacks on the MNIST dataset, the proposed method achieved a low 2% success rate for identity reconstruction, compared to 18% for literature [21] and 22% for literature [22], which employ less advanced differential privacy techniques. This low success rate demonstrates RAPPOR's superior noise injection strategy.