Multi-modal adversarial optimization engine and deployment integrator

A computer-implemented method for modulating detectability of a target object by machine perception, comprising:

• receiving input data including a representation of the target object and an objective corresponding to reduce detectability or induce a designated detection of the target object;
• generating, using an adversarial pattern generator, a plurality of candidate adversarial patterns based on the objective for the target object;
• simulating, for each candidate adversarial pattern of the plurality of adversarial patterns, an integration of the target object in a virtual environment across a plurality of sensor modalities;
• evaluating each simulated candidate adversarial pattern using an ensemble of machine-learned detector models corresponding to the plurality of sensor modalities, the detector models producing detection outputs corresponding to the target object;
• calculating a performance metric for each simulated candidate adversarial pattern based on the detection outputs, the performance metric corresponding to the objective;
• determining whether the performance metric reaches a predetermined threshold level for the objective;
• and
• performing one of:
• selecting the simulated adversarial pattern as an optimized adversarial pattern in response to reaching the performance metric, or
• iteratively adjusting, in response to not reaching the predetermined threshold level performance metric, the simulated candidate adversarial pattern until it converges on meeting the predetermined threshold level metric for an optimized output adversarial pattern for the target object;
• and
• compiling the optimized adversarial pattern into an output file for the target object.

A system to influence detectability of a target object, the system comprising:

• an input processing module configured to receive data of a target object and context parameters corresponding to detectability of the target object;
• a pattern generator module configured to generate a plurality of candidate adversarial patterns for the target object;
• a multi-modal simulation module configured to:
• simulate the target object with each candidate adversarial pattern of the plurality of candidate adversarial patterns, the simulation in a virtual environment across a plurality of sensor modalities, and,
• produce simulated sensor outputs for each sensor modality of the plurality of sensor modalities;
• a detector ensemble module comprising a plurality of trained detector models each corresponding to a sensor modality of the plurality of sensor modalities, the detector ensemble module configured to process the simulated sensor outputs to produce detection metrics indicative of whether the target object would be detected or misclassified under each candidate adversarial pattern of the plurality of candidate adversarial patterns;
• an optimization module configured to iteratively adjust the candidate adversarial patterns until a threshold is reached at which point an optimized adversarial pattern is generated;
• and
• an output compiler module configured to generate output data for the optimized adversarial pattern.