🔮 Nextcloud Suspicious Login Detection

Detect and warn about suspicious IPs logging into Nextcloud

The app is still under development, so it’s time for you to get involved! 👩‍💻

How it works

Data collection

Once this app is enabled, it will automatically start tracking (IP, uid) tuples from successful logins on the instance and feed them into the login_address table. This insert operation is executed for the majority of requests (client authenticate on almost all requests) and therefore has to be fast. In a background job, these rows will be transformed into an aggregated format that is suitable for the training of the neural net. The (IP, uid) tuple becomes (IP, uid, first_seen, last_seen, seen) so that we know which (IP, uid) tuple has been seen first and last. The aggregated data is a compressed format of the raw data. The original data gets deleted and thus the database does not need much space for the collected login data.

Neural net

When enough data is collected – which by default is 60 days (!) – a first training run can be started.

The app registers a background job that invokes the training once a day. As long as there isn't sufficient data, no trained model is generated.

Manual training

The training can also be invoked via the OCC command line tool:

php -f occ suspiciouslogin:train

This command uses several sensible default that should work for instances of any size. The --stats flag is useful to see the measured performance of the trained model after the training finishes. The duration of the training run depends on the size of the input training set, but is usually between two to 15 minutes.

The full list of parameters, their description and default values can be seen with

php -f occ suspiciouslogin:train --help

Hyper parameter optimization (optional)

To find the best possible parameters for the training it's possible to start a hyper parameter optimization run via the CLI:

php -f occ suspiciouslogin:optimize

This command uses the heuristic simulated annealing algorithm to find optimal parameter sets in the multidimensional parameter space. By default this will do 100 steps consisting of five training runs per step, hence this command might take a few days to execute on large instances. On smaller ones it will also take a few hours.

Login classification

As soon as the first model is trained, the app will start classifying (IP, uid) tuples on login. In contrast to the data collection it won't consider requests authenticated via an app password as suspicious. Should it detect a password login where the (IP, uid) is classified as suspicious by the trained model, it will add an entry to the suspicious_login table, including the timestamp, request id and URL.

Development setup

1. ☁ Clone the app into the apps folder of your Nextcloud: git clone https://github.com/nextcloud/suspicious_login.git
2. 💻 Run npm i or krankerl up to install the dependencies
3. 🏗 To build the Javascript whenever you make changes, run npm run dev
4. ☁ Enable the app through the app management of your Nextcloud or run krankerl enable
5. 👍 Partytime! Help fix some issues and review pull requests