ML-Predicted Coverage Feedback for Black Box Fuzzing

The patent introduces an intermediary component (the machine learning model) that mediates between the black box fuzzer and the target program. This model learns to predict code coverage and other feedback information by analyzing the relationship between inputs and outputs, effectively creating a virtual feedback channel without requiring direct access to the target program's internal state or source code.

The patent creates a copy or approximation of the target program's internal behavior through a trained machine learning model. This model replicates the coverage information that would otherwise be available only through direct access to the target program, allowing the fuzzer to operate with feedback as if it had gray box capabilities while maintaining black box operation.

The patent implements a feedback mechanism where the machine learning model predicts coverage information based on input characteristics, and this predicted feedback is used to guide the generation of subsequent test cases. The fuzzer uses this feedback to prioritize inputs that are likely to explore new code paths, transforming random testing into a directed search that improves over time without requiring access to target program internals.

The patent performs preliminary training of the machine learning model using a dataset of inputs and their corresponding coverage information. This preliminary action equips the model with learned patterns that enable it to predict coverage for new inputs, allowing the fuzzer to make informed decisions from the start rather than testing purely at random.