Kui: AI-Powered Anti-Spam System Architecture and Strategies

This article introduces Kui (钟馗), an AI-powered anti-spam system developed by iQiyi's technical team to combat malicious activities in online communities. The system addresses various forms of spam including advertising, pornography, and harmful content that threaten community stability.

The system architecture follows a three-layer design: service layer providing HTTP interfaces for business partners, algorithm strategy layer containing preprocessing, rule-based, keyword, risk control, and model modules, and auxiliary modules including human review systems, corpus annotation, ML model training, log search, and monitoring systems.

Key strategies include:

1. Keyword Strategy (七伤拳): Uses various matching algorithms including hard matching, jump-word matching, contact information matching, pinyin matching, punctuation matching, long sentence fuzzy matching, and exact matching.

2. Rule Strategy (独孤九剑): Employs dozens of rules for detecting spam, including longest continuous letter detection, sentiment analysis using dependency parsers, and other heuristic-based approaches.

3. Model Strategy (九阴真经): Implements multiple machine learning models including Logistic Regression, LSTM, Convolutional-LSTM, CNN, BiLSTM-Attention, and cw2vec-Attention. These models handle various text types including pornographic, advertising, meaningless, abusive, and vulgar content.

4. Risk Control Strategy (葵花点穴手): Applies user-level and device-level rules to control spam behavior, such as detecting posting frequency and similarity patterns.

The article emphasizes that anti-spam is an ongoing battle, with black market operators continuously evolving their techniques including group control, cloud control, and AI technologies. The system must continuously improve and adapt to maintain effectiveness.