Systems and Methods With Robust Classifiers That Defend Against Patch Attacks

A computer-implemented method for generating machine learning predictions with defenses against patch attacks, the method comprising:

• obtaining sensor data from a sensor system that includes at least one sensor;
• obtaining a first digital image based on the sensor data;
• generating a first set of location data via a random process, the first set of location data including a random selection of locations with respect to the first digital image;
• cropping the first digital image with respect to each location data of the first set of location data such that a first set of random crops is generated based on the first set of location data in which each crop of the first set of random crops is a different region of the first digital image;
• generating, via a machine learning system, a respective prediction for each crop of the first set of random crops to create a first set of predictions;
• evaluating the first set of predictions collectively to determine a majority prediction from among the first set of predictions;
• and
• generating an output label that identifies the first digital image as possessing the majority prediction.

A system for performing classifications with defenses against patch attacks, the system comprising:

• a sensor system including at least one sensor to capture sensor data;
• one or more non-transitory computer readable storage devices including a machine learning system and a framework configured to defend against patch attacks;
• and
• a processing system including at least one computer processor that is communicatively connected at least to the one or more non-transitory computer readable storage devices, the processing system being configured to execute a method corresponding to the framework that includes:
• obtaining a first digital image based on the sensor data;
• generating a first set of location data via a random process, the first set of location data including a random selection of locations with respect to the first digital image;
• cropping the first digital image with respect to each location data of the first set of location data such that a first set of random crops is generated based on the first set of location data where each crop is a different region of the first digital image;
• classifying, via the machine learning system, each crop of the first set of random crops to create a first set of class predictions;
• evaluating the first set of class predictions collectively to determine a majority class from among the first set of class predictions;
• and
• generating a class label that identifies the first digital image as belonging to the majority class.