Black Box Fuzzing Feedback via Machine Learning Prediction

The patent implements feedback mechanisms in black box fuzzing by using machine learning models to predict code coverage and path coverage metrics. The fuzzer collects actual coverage data from executed test cases, uses this data to train or fine-tune the ML model, and then employs the model's predictions to guide subsequent test case generation. This creates a closed-loop feedback system where coverage information from previous iterations informs future fuzzing directions, enabling continuous improvement without requiring source code modification or access during the fuzzing process itself

The patent applies preliminary action by pre-training machine learning models on available training data that includes information about the target program's behavior and coverage characteristics. This pre-training phase prepares the ML model to make accurate predictions about code coverage and path coverage before the actual fuzzing begins. The model is ready to provide guidance from the start, eliminating the need for iterative model development during fuzzing and enabling immediate directed test case generation

The patent introduces machine learning models as intermediary components between the fuzzer and the target program. These ML models act as mediators that translate raw coverage data into predictive insights about which test cases are likely to uncover vulnerabilities. The intermediary model layer processes coverage information and generates predictions about code coverage and path coverage, guiding the fuzzer to generate more targeted test cases without requiring direct access to source code or complex instrumentation of the target program

The patent replaces traditional mechanical feedback mechanisms (which require source code instrumentation, breakpoints, and direct program-fuzzer communication) with a data-driven machine learning approach. Instead of using complex mechanical systems to track and report coverage information, the patent uses ML models trained on coverage data to predict which test cases will be effective. This substitution reduces the mechanical complexity of the fuzzing system while maintaining or improving targeting accuracy

The patent implements feedback mechanisms that work specifically for embedded systems with non-deterministic outputs by using machine learning models to capture and predict coverage patterns. The system collects coverage data from executed test cases on the embedded system, trains ML models on this data to learn the system's behavior patterns, and then uses these models to guide subsequent test case generation. This feedback loop adapts to the non-deterministic nature of embedded systems by learning from actual execution patterns rather than relying on deterministic source code analysis

The patent applies parameter changes by using machine learning models to dynamically adjust the parameters and characteristics of generated test cases based on predicted coverage information. The ML models analyze coverage patterns and modify test case parameters (such as input data characteristics, execution paths, or test conditions) to target areas of the embedded system that are less frequently exercised. This dynamic parameter adjustment enables effective fuzzing of non-deterministic embedded systems by adapting to their specific behavior patterns