Robot Security Anomaly Detectionmark

Robot Security Anomaly Detection Datasets

Stefano Simonazzi, Lorenzo Brigato, Riccardo Sartea, Christian Napoli, Alessandro Farinelli and Luca Iocchi

DIAG - Sapienza University of Rome, Italy

Dpt. of Computer Science, University of Verona, Italy

This web site contains data sets for Robot Security Anomaly Detection

Datasets download

Boat

Pepper

Boat

Autonomous surface vehicles (ASVs) can be used for persistent large-scale monitoring of aquatic environments. ASVs are a valid alternative to the traditional manual sampling approach since they are capable of undertaking long-endurance missions collecting water quality data.

Low-cost Autonomous surface vehicles (ASVs) are employed in the EU-funded project INTCATCH to develop efficient and user-friendly monitoring strategies for facilitating sustainable water quality management by non-experts.

This system, like all autonomous drones, is subject to cyber-attacks and hacking. Moreover, ASVs for water monitoring transmit the acquired data regarding water quality (e.g. Dissolved Oxygen, Ph level, Electrical Conductivity and so on) using standard networks (a WiFi network and the 3G mobile network in our case), hence such data could be exposed to tampering.

The dataset has 41 variables: 27 reals, 8 integers, 6 strings of which 2 IP address.

The attack dataset contains the following simulated attacks:

DosPay: (27%) DOS attack with 32 bytes payload. A DOS attack was simulated during a normal monitoring session by using Kali Linux and hping3 tool. The size of the payload is a random value of size 32 byte. The attack had success and the communication between the server and client application become unavailable;
Dos: (19%) DOS attack with no payload. This attack is similar to the previous one, but with the difference that the payload was empty. The small packet size allows a higher sending frequency which causes a faster connection interruption;
GpsDown: (27%) GPS down fault. This simply simulates the GPS interruption;
Stuck: (27%) Vehicle got stuck fault. This simulates the case when the boat got stuck somewhere. The boats are designed to be slim and agile in order to run in small water areas, but when they monitor along a shore they could get caught up and get stuck due to rubbish or simple vegetation.

Pepper

SoftBank Pepper is a robot designed for human-robot interaction and social robotics and it is commonly used in public spaces to interact with people.

The attacks are simulated taking partial control of some devices of the robot, namely the joints of the arms and the body, the LEDs, and the wheels, by connecting to the NaoQi system on port 9559 without authentication.

The dataset has 256 real values from different sensors, corresponding to: joints positions, wheels velocity, laser range, sonar range, led configuration, gyroscope, accelerometers, touch sensors, motors stiffnesses, motors temperatures.

The dataset with only normal records (taken from Pepper's built-in autonomous interactive behavior) has been collected for training, while the attack dataset contains the following simulated attacks:

LedsControl: The robot's LEDs are controlled by a remote program and activated in a pseudo-random mode;
JointsControl: The robot's joints are controlled by a remote program setting joints values to obtain special poses of the robot;
WheelsControl: The robot movement is controlled by an external program connected to a remote joystick to navigate in the environment.

One-class Anomaly Detection

The proposed datasets are divided into normal behavior and attacks and can be used as required in your work.

In our experimentation, we followed a one-class approach to anomaly detection. We selected a portion of normal data for the model's training and used the rest together with the attacks for test. In detail, we considered 2,200 normal training records for the boat and the remaining 938 for the test, while for pepper we considered 14,000 normal training records and the remaining 4,244 for the test. In our work the models used, being temporal, require in input a sequence of records in chronological order and contiguous. For this reason, it is necessary to select the normal test records from the total of normal records by keeping the order. Test data containing n normal records are created by randomly selecting 10 contiguous sequences of n/10 records from the normal data, the remaining records are used for training.

Report abuse