Sharding technology is considered the most promising solution to address the scalability issues of blockchain. Most existing sharding techniques focus on performance improvement but fail to consider the categories of nodes when assigning them to shards. This may lead to malicious clustering in the same shard, reducing the security of the sharded blockchain. This paper introduces the random forest model for classifying nodes and proposes the ANC (Assign Nodes by Category) algorithm, which isolates malicious nodes and distributes the remaining nodes to different shards. During the process of transaction execution, we monitor nodes to detect whether malicious behavior has occurred, enhancing system security. Experimental results demonstrate that the node classification result achieves 97% accuracy and indirectly increases the safety of Blockchain. Compared to the baseline models, the proposed solution is superior in terms of throughput, latency, and security.