behaviorIntention

introduction

      Most of the sequence behavior log analysis and anomaly identification tools currently on the market are designed for specific systems, and the identification rules are usually only applicable to specific environments. These tools often require redesigning and adjusting analysis models for different systems and lack universality. Even for machine learning-based solutions, behavioral log data from different systems still need to be manually labeled, which increases the complexity and cost of operations and is difficult to be universally applicable to multiple systems. In response to this situation, the R&D team of this project proposed a digital security model based on deep learning. The model automatically mines the potential system behavior rules and operation logic by learning training data composed entirely of normal sequence behavior log data, thereby realizing the recognition of sequence behavior and intention of behavior logs. In addition, when it is impossible to obtain samples consisting entirely of normal log data, the model parameters can be adjusted to use mixed data consisting mainly of normal behavior sequence log data (where normal logs account for more than 90%) for model training. Compared with traditional methods, this model does not need to rely on manual labeling, has high intelligence and accuracy, has wide adaptability and versatility, and can provide efficient security protection for various information systems.

Overall model architecture