Applications change frequently as new URLs are added and existing parameters provide new functions. This means the mathematical model of the same parameter might be different from what FortiWeb originally observed during the collection phase. In this case, FortiWeb needs to re-learn the parameter and updates the mathematical model for it.

Enable this option to automatically update the mathematical models of the parameters when they are changed.

This option appears when you enable Dynamically update when parameters change.

The default value is 2, which means if 2 newly generated boxplots don't overlap with any one of the sample boxplots, FortiWeb automatically updates the machine learning model.

You can set a value from 1 to 3.

This option appears when you enable Dynamically update when parameters change.

The system uses boxplots to determine whether a parameter has changed. The boxplot displays the probability distribution of the parameter value. During sample collection period, the system generates 2 or 4 boxplots. After machine learning model is built, the system will keep on generating new boxplots to display the probability distribution of the new inputs. If the probability distribution area of the newly generated boxplot doesn't overlap with any one of the sample boxplots, the system determines this parameter has changed.

Depending on the Application Change Sensitivity level, the system triggers model update when it observes different extent of overlapping area.

• Low—The system triggers model update only when the entire data distribution area (from the maximum value to the minimum value, that is, the entire area containing all the data) of the new boxplot doesn't have any overlapping part with that of the sample boxplots.
• Medium—The system triggers model update if the notch area (the median rectangular area in the boxplot where most of the data is located) of the new boxplot doesn't have any overlapping part with the entire data distribution areas of the sample boxplots.
• High—The system triggers model update as long as the notch area of the new boxplot doesn't have any overlapping part with that of the sample boxplots.

There are two anomaly detection methods: Automatic and Quantile. They determine anomalies with the aid of machine learning models. The machine learning model judges whether a request is normal or not based on its HMM probability and the length of the parameter value.

Both methods have potential anomaly threshold and definite anomaly threshold to detect anomalies. You can set different actions (alert, alert&deny or period block) for potential anomaly and definite anomaly.

Automatic

Automatic is the default method. Compared with Quantile method, Automatic method is more complicated, so that it requires longer time to detect anomalies, but it's more accurate.

The value for the strictness level can range from 0.1 to 1.0. The higher the value, the more anomalies will be triggered. For example, 0.1 means that 0.1% of all samples with the largest HMM probability and length will be treated as anomalies.

Quantile

Quantile simply uses a threshold, and all probabilities above it are identified as anomalies.

The value for the strictness level can range from 0.1 to 1.0. The higher the value, the more anomalies will be triggered. If you select 0.3 for Potential Anomaly, it means that the 99.7 quantile will be used as the threshold. If a probability exceeds this 99.7 quantile threshold, it will be considered to be an unexpected outlier, so the request would be considered as an attack.

Enter the threshold value or choose the threshold numbers. The default is 0.3, and the valid range is from 0 to 1. The higher the threshold, the more anomalies will be triggered.

Enter the threshold value or choose threshold numbers. The default is 0.1, and the valid range is from 0 to 0.9. The higher the threshold, the more anomalies will be triggered.

Enable to scan anomalies to verify whether they are attacks. It provides a method to check whether an anomaly is a real attack by the trained Support Vector Machine Model.

Click the View Threat Models link to do more detailed configuration for your “Threat Model”. Currently, seven trained Support Vector Machine Model are provided for seven attack types. You may choose to enable or disable them individually.

Note:

All requests are scanned first by HMM and then by Threat model. Choose the action FortiWeb takes when attack is verified for each of the following situations:

• Alert—Accepts the connection and generate an alert email and/or log message.
• Alert & Deny—Blocks the request (or reset the connection) and generates an alert and/or log message.

Select an action FortiWeb will take when machine learning detects a potential anomaly. The default value is Alert.

Select an action FortiWeb will take when machine learning detects a definite anomaly. The default value is Alert & Deny.

Select an action the FortiWeb appliance will take when machine learning detects an HTTP Method Violation. The default value is Alert & Deny.