As a distributed ledger technology, blockchain has excellent openness and transparency, which can provide data security management services for distributed intelligent systems and establish effective security guarantee mechanisms. However, precisely due to the open nature of blockchain, malicious users can trace the real transaction transfer path with high probability and even obtain the real identity of users by collecting transaction information on the blockchain and performing data analysis. Besides, existing intelligent systems lack effective encryption or desensitization measures, and attackers are able to access data in intelligent systems through identity forgery and other means, posing a huge potential risk to user privacy. To alleviate the above security issues, this paper proposes an efficient blockchain-based privacy-preserving scheme with attribute and homomorphic encryption, which can not only achieve user-level fine-grained secure access control but also supports the transmission and verification of blockchain data in the form of ciphertext. The access control method based on attribute-based encryption and the data transmission method based on homomorphic encryption are proposed, and the blockchain-based access whitelist mechanism is designed to reduce the resource loss due to repeated authentication. Simulation calculations and blockchain performance experiments show that this scheme can develop flexible access policies according to the attributes of users, has good performance in computational efficiency, and the blockchain performance test results are stable with errors in milliseconds for all operations. It has certain application potential in the field of distributed intelligent systems and blockchain.

The prosperity of social networks provides users with convenient communication but also attracts a large number of spammers. To solve this problem, this paper combines supervised learning and unsupervised learning algorithms, and proposes a novel hybrid model based on OPTICS and SVM. First, we collected a dataset from Sina Weibo including 10,000 users and 134,188 messages; then extracted the content based features and user behavior based features from the dataset; afterwards, we applied the features into the hybrid model to establish the classification model. The experiment shows that the proposed approach is capable of detecting spammers effectively with 87.6% spammers and 94.7% legitimate users correctly classified.

A social spammer detection model based on tri-training (SSDTT) is adopted. The main procedure of the work is: First, train three original classifiers with a small amount of labeled data. Then, select confident users that are labeled for a classifier if the other two classifiers agree on the labeling as new training data. Afterwards, repeat these steps until three classifiers are not updated. Experimental results indicate that SSDTT has the same performance with the supervised learning in the case of lacking sufficient labeled data.