Year: 2026
Type: article
Abstract: In the research of artificial intelligence (AI) for social network security, in response to the difficulties in engineering implementation caused by the disconnection between models and systems and the lack of standards for risk assessment, this paper constructs a risk assessment framework and a modular scalable engineering system that integrates multidimensional standard indicators. The system decomposes social network security risks into four categories: data poisoning, model theft, adversarial attacks, and privacy breaches, formulated as a four-dimensional multi-label risk state whose dimensions are anchored to distinct security properties of the AI lifecycle, and maps them to quantifiable sub indicators through explicit functional mapping, with cross-dimension co-occurrence retained in the annotation and measured empirically. This article designs a streaming processing pipeline consisting of data access, feature extraction, risk inference, and response generation, and implements micro-batch event stream processing through Kafka. It integrates a federated learning mechanism based on Secure Aggregation, supports local model training and encryption parameter aggregation, and introduces a bias triggering strategy to reduce communication overhead, and composes a Gaussian perturbation step with the masked aggregation so that the released aggregate carries an explicit Rényi differential privacy bound. This article develops an automated response orchestration engine that interfaces with the platform’s security control plane and routes irreversible disposal actions to human operators under a versioned, reversible audit log. The experimental results demonstrate that in typical social platform scenarios, the proposed system attains accuracy with a mean of 0.91 to 0.94 and F1 score with a mean of 0.89 to 0.92, reflecting consistent cross scenario risk identification performance. Under a sustained input load of 20,000 events per second, the 99th percentile of the end-to-end latency measured from event ingestion to response acknowledgement is 310 ms and the sustainable throughput is 28,500 events per second. Under no privacy conditions, the success rate of member inference attacks is 18.2%, with an accuracy decrease of no more than 2.3%, indicating good model utility retention ability. The source code, the experiment configuration files, and a synthetic data generator reproducing the anonymized event schema are released at https://doi.org/10.5281/zenodo.22272217 .
Source: Scientific Reports
Authors: Dongwei Kang, Lingli Qing, Chao He, Ying Chen
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7225897931
