Every account, user, service, and machine has a digital fingerprint that represents the typical actions performed and not performed over a given period of time. Understanding every entity's day-to-day, moment-by-moment work helps us identify anomalous behavior and uncover potential threats in the environment.
To construct this digital fingerprint, we will be training unsupervised behavioral models at various granularities, including a generic model for all users in the organization along with fine-grained models for each user to monitor their behavior. These models are continuously updated and retrained over time, and alerts are triggered when deviations from normality occur for any user.
The data we will want to use for the training and inference will be any sensitive system that the user interacts with, such as VPN, authentication and cloud services. The digital fingerprinting example (examples/digital_fingerprinting/README.md) included in Morpheus ingests logs from AWS CloudTrail, Azure Active Directory, and Duo Authentication.
The location of these logs could be either local to the machine running Morpheus, a shared file system like NFS, or on a remote store such as Amazon S3.
Additional data sources and remote stores can easily be added using the Morpheus API. The key to applying DFP to a new data source is through the process of feature selection. Any data source can be fed into DFP after some preprocessing to get a feature vector per log/data point. In order to build a targeted model for each entity (such as user, service and machine), the chosen data source needs a field that uniquely identifies the entity we're trying to model.
Adding a new source for the DFP pipeline requires defining five critical pieces:
-
The user ID column specified by the
morpheus.config.Config.ae.userid_column_nameconfiguration attribute. This can be any column which uniquely identifies the user, account or service being fingerprinted. Examples of possible user IDs could be:- A username or full name (for example,
"johndoe"or"Jane Doe") - User's LDAP ID number
- A user group (for example,
"sales"or"engineering") - Host name of a machine on the network
- IP address of a client
- Name of a service (for example,
"DNS","Customer DB", or"SMTP")
- A username or full name (for example,
-
The timestamp column specified by the
morpheus.config.Config.ae.timestamp_column_nameconfiguration attribute. The timestamp column needs to be converted to adatetimecolumn refer toDateTimeColumn. -
The model's features as a list of strings specified by the
morpheus.config.Config.ae.feature_columnsconfiguration attribute, which should all be available to the pipeline after theDFPPreprocessingStage. -
A
DataFrameInputSchemafor theDFPFileToDataFrameStagestage. -
A
DataFrameInputSchemafor theDFPPreprocessingStage.
The DFP workflow is provided as two separate examples: a simple, "starter" pipeline for new users and a complex, "production" pipeline for full scale deployments. While these two examples both perform the same general tasks, they do so in very different ways. The following is a breakdown of the differences between the two examples.
This example is designed to simplify the number of stages and components and provide a fully contained workflow in a single pipeline.
Key Differences:
- A single pipeline which performs both training and inference
- Requires no external services
- Can be run from the Morpheus CLI
This example is described in more detail in examples/digital_fingerprinting/starter/README.md.
This example is designed to illustrate a full-scale, production-ready, DFP deployment in Morpheus. It contains all of the necessary components (such as a model store), to allow multiple Morpheus pipelines to communicate at a scale that can handle the workload of an entire company.
Key Differences:
- Multiple pipelines are specialized to perform either training or inference
- Requires setting up a model store to allow the training and inference pipelines to communicate
- Organized into a docker-compose deployment for easy startup
- Contains a Jupyter notebook service to ease development and debugging
- Can be deployed to Kubernetes using provided Helm charts
- Uses many customized stages to maximize performance.
This example is described in examples/digital_fingerprinting/production/README.md as well as the rest of this document.
| Feature | Description |
|---|---|
userIdentityaccessKeyId |
for example, ACPOSBUM5JG5BOW7B2TR, ABTHWOIIC0L5POZJM2FF, AYI2CM8JC3NCFM4VMMB4 |
userAgent |
for example, Mozilla/5.0 (compatible; MSIE 9.0; Windows NT 10.0; Trident/5.1), Mozilla/5.0 (Linux; Android 4.3.1) AppleWebKit/536.1 (KHTML, like Gecko) Chrome/62.0.822.0 Safari/536.1, Mozilla/5.0 (Macintosh; U; PPC Mac OS X 10 7_0; rv:1.9.4.20) Gecko/2012-06-10 12:09:43 Firefox/3.8 |
userIdentitysessionContextsessionIssueruserName |
for example, role-g |
sourceIPAddress |
for example, 208.49.113.40, 123.79.131.26, 128.170.173.123 |
userIdentityaccountId |
for example, Account-123456789 |
errorMessage |
for example, The input fails to satisfy the constraints specified by an AWS service., The specified subnet cannot be found in the VPN with which the Client VPN endpoint is associated., Your account is currently blocked. Contact aws-verification@amazon.com if you have questions. |
userIdentitytype |
for example, FederatedUser |
eventName |
for example, GetSendQuota, ListTagsForResource, DescribeManagedPrefixLists |
userIdentityprincipalId |
for example, 39c71b3a-ad54-4c28-916b-3da010b92564, 0baf594e-28c1-46cf-b261-f60b4c4790d1, 7f8a985f-df3b-4c5c-92c0-e8bffd68abbf |
errorCode |
for example, success, MissingAction, ValidationError |
eventSource |
for example, lopez-byrd.info, robinson.com, lin.com |
userIdentityarn |
for example, arn:aws:4a40df8e-c56a-4e6c-acff-f24eebbc4512, arn:aws:573fd2d9-4345-487a-9673-87de888e4e10, arn:aws:c8c23266-13bb-4d89-bce9-a6eef8989214 |
apiVersion |
for example, 1984-11-26, 1990-05-27, 2001-06-09 |
| Feature | Description |
|---|---|
appDisplayName |
for example, Windows sign in, MS Teams, Office 365 |
clientAppUsed |
for example, IMAP4, Browser |
deviceDetail.displayName |
for example, username-LT |
deviceDetail.browser |
for example, EDGE 98.0.xyz, Chrome 98.0.xyz |
deviceDetail.operatingSystem |
for example, Linux, IOS 15, Windows 10 |
statusfailureReason |
for example, external security challenge not satisfied, error validating credentials |
riskEventTypesv2 |
AzureADThreatIntel, unfamiliarFeatures |
location.countryOrRegion |
country or region name |
location.city |
city name |
| Feature | Description |
|---|---|
logcount |
tracks the number of logs generated by a user within that day (increments with every log) |
locincrement |
increments every time we observe a new city (location.city) in a user's logs within that day |
appincrement |
increments every time we observe a new app (appDisplayName) in a user's logs within that day |
| Feature | Description |
|---|---|
auth_device.name |
phone number |
access_device.browser |
for example, Edge, Chrome, Chrome Mobile |
access_device.os |
for example, Android, Windows |
result |
SUCCESS or FAILURE |
reason |
reason for the results, for example, User Cancelled, User Approved, User Mistake, No Response |
access_device.location.city |
city name |
| Feature | Description |
|---|---|
logcount |
tracks the number of logs generated by a user within that day (increments with every log) |
locincrement |
increments every time we observe a new city (location.city) in a user's logs within that day |
DFP in Morpheus is accomplished via two independent pipelines: training and inference. The pipelines communicate via a shared model store (MLflow), and both share many common components, as Morpheus is composed of reusable stages that can be easily mixed and matched.
- Trains user models and uploads to the model store
- Capable of training individual user models or a fallback generic model for all users
- Downloads user models from the model store
- Generates anomaly scores per log
- Sends detected anomalies to monitoring services
- Detected anomalies are published to an S3 bucket, directory or a Kafka topic.
- Output can be integrated with a monitoring tool.
DFP in Morpheus is built as an application of containerized services and can be run in two ways:
- Using docker-compose for testing and development
- Using helm charts for production Kubernetes deployment
The reference architecture is composed of the following services:
| Service | Description |
|---|---|
mlflow |
MLflow provides a versioned model store |
jupyter |
Jupyter Server necessary for testing and development of the pipelines |
morpheus_pipeline |
Used for executing both training and inference pipelines |
- Docker and docker-compose installed on the host machine
- Supported GPU with NVIDIA Container Toolkit
Note: For GPU Requirements refer to the Getting Started guide.
From the root of the Morpheus repo, run:
cd examples/digital_fingerprinting/production
docker compose buildNote: This requires version 1.28.0 or higher of Docker Compose, and preferably v2. If you encounter an error similar to:
ERROR: The Compose file './docker-compose.yml' is invalid because: services.jupyter.deploy.resources.reservations value Additional properties are not allowed ('devices' was unexpected)This is most likely due to using an older version of the
docker-composecommand, instead re-run the build withdocker compose. Refer to Migrate to Compose V2 for more information.
First, we will need to install s3fs and then run the examples/digital_fingerprinting/fetch_example_data.py script. This will download the example data into the examples/data/dfp dir.
From the Morpheus repo, run:
pip install s3fs
python examples/digital_fingerprinting/fetch_example_data.py allFrom the examples/digital_fingerprinting/production dir run:
docker compose up jupyterOnce the build is complete and the service has started, a message similar to the following should display:
jupyter | To access the server, open this file in a browser:
jupyter | file:///root/.local/share/jupyter/runtime/jpserver-7-open.html
jupyter | Or copy and paste one of these URLs:
jupyter | http://localhost:8888/lab?token=<token>
jupyter | or http://127.0.0.1:8888/lab?token=<token>
Copy and paste the URL into a web browser. There are four notebooks included with the DFP example:
- dfp_azure_training.ipynb - Training pipeline for Azure Active Directory data
- dfp_azure_inference.ipynb - Inference pipeline for Azure Active Directory data
- dfp_duo_training.ipynb - Training pipeline for Duo Authentication
- dfp_duo_inference.ipynb - Inference pipeline for Duo Authentication
Note: The token in the URL is a one-time use token and a new one is generated with each invocation.
By default, the morpheus_pipeline will run the training pipeline for Duo data from the examples/digital_fingerprinting/production dir run:
docker compose up morpheus_pipelineIf instead you want to run a different pipeline from the examples/digital_fingerprinting/production dir, run:
docker compose run morpheus_pipeline bashFrom the prompt within the morpheus_pipeline container, you can run either the dfp_azure_pipeline.py or dfp_duo_pipeline.py pipeline scripts.
python dfp_azure_pipeline.py --help
python dfp_duo_pipeline.py --helpBoth scripts are capable of running either a training or inference pipeline for their respective data sources. The command-line options for both are the same:
| Flag | Type | Description |
|---|---|---|
--train_users |
One of: all, generic, individual, none |
Indicates whether or not to train per user or a generic model for all users. Selecting none runs the inference pipeline. |
--skip_user |
TEXT | User IDs to skip. Mutually exclusive with only_user |
--only_user |
TEXT | Only users specified by this option will be included. Mutually exclusive with skip_user |
--start_time |
TEXT | The start of the time window, if undefined start_date will be now()-duration |
--duration |
TEXT | The duration to run starting from start_time [default: 60d] |
--cache_dir |
TEXT | The location to cache data such as S3 downloads and pre-processed data [environment variable: DFP_CACHE_DIR; default: ./.cache/dfp] |
--log_level |
One of: CRITICAL, FATAL, ERROR, WARN, WARNING, INFO, DEBUG |
Specify the logging level to use. [default: WARNING] |
--sample_rate_s |
INTEGER | Minimum time step, in milliseconds, between object logs. [environment variable: DFP_SAMPLE_RATE_S; default: 0] |
-f, --input_file |
TEXT | List of files to process. Can specify multiple arguments for multiple files. Also accepts glob (*) wildcards and schema prefixes such as s3://. For example, to make a local cache of an s3 bucket, use filecache::s3://mybucket/*. Refer to fsspec documentation for list of possible options. |
--watch_inputs |
FLAG | Instructs the pipeline to continuously check the paths specified by --input_file for new files. This assumes that the at least one paths contains a wildcard. |
--watch_interval |
FLOAT | Amount of time, in seconds, to wait between checks for new files. Only used if --watch_inputs is set. [default 1.0] |
--tracking_uri |
TEXT | The MLflow tracking URI to connect to. [default: http://localhost:5000] |
--help |
Show this message and exit. |
To run the DFP pipelines with the example datasets within the container, run:
-
Duo Training Pipeline
python dfp_duo_pipeline.py --train_users=all --start_time="2022-08-01" --input_file="/workspace/examples/data/dfp/duo-training-data/*.json"
-
Duo Inference Pipeline
python dfp_duo_pipeline.py --train_users=none --start_time="2022-08-30" --input_file="/workspace/examples/data/dfp/duo-inference-data/*.json"
-
Azure Training Pipeline
python dfp_azure_pipeline.py --train_users=all --start_time="2022-08-01" --input_file="/workspace/examples/data/dfp/azure-training-data/*.json"
-
Azure Inference Pipeline
python dfp_azure_pipeline.py --train_users=none --start_time="2022-08-30" --input_file="/workspace/examples/data/dfp/azure-inference-data/*.json"
The output files will contain those logs from the input dataset for which an anomaly was detected; this is determined by the z-score in the mean_abs_z field. By default, any logs with a z-score of 2.0 or higher are considered anomalous. Refer to DFPPostprocessingStage.
Most of the fields in the output files generated by running the above examples are input fields or derived from input fields. The additional output fields are:
| Field | Type | Description |
|---|---|---|
event_time |
TEXT | ISO 8601 formatted date string, the time the anomaly was detected by Morpheus |
model_version |
TEXT | Name and version of the model used to performed the inference, in the form of <model name>:<version> |
max_abs_z |
FLOAT | Max z-score across all features |
mean_abs_z |
FLOAT | Average z-score across all features |
In addition to this, for each input feature the following output fields will exist:
| Field | Type | Description |
|---|---|---|
<feature name>_loss |
FLOAT | The loss |
<feature name>_z_loss |
FLOAT | The loss z-score |
<feature name>_pred |
FLOAT | The predicted value |
Refer to DFPInferenceStage for more on these fields.
Starting the morpheus_pipeline or the jupyter service, will start the mlflow service in the background. For debugging purposes, it can be helpful to view the logs of the running MLflow service.
From the examples/digital_fingerprinting/production dir, run:
docker compose up mlflow- Kubernetes cluster configured with GPU resources
- NVIDIA GPU Operator installed in the cluster
Note: For GPU Requirements refer to the Getting Started guide.
For details on customizing the DFP pipeline refer to Digital Fingerprinting (DFP) Reference.