Modern Cyber-Physical Systems (CPS) increasingly utilize AI-enabled controllers with deep neural networks (DNNs). As these systems grow in complexity, traditional verification methods become less effective. A practical alternative is \emph{falsification}, a search for inputs that cause the system to violate specified properties. Falsification techniques can be categorized into \emph{black-box} methods, which are implementation agnostic and rely on general heuristics to guide the search, and \emph{white-box} methods, which utilize implementation-specific information to better direct the search, but their applicability may be limited to specific controller types.
This paper introduces an input-model-driven falsification approach, a novel black-box technique that leverages a learned representation of the input domain for better directing the search without relying on the internals of the system model being falsified. By utilizing a Variational Autoencoder (VAE) to encode inputs into low-dimensional embeddings, this approach captures useful problem-domain-specific information to guide the falsification process without sacrificing the generality of black-box techniques. Evaluations on CPS models from the autonomous driving domain show a falsification success rate of 60.18%, outperforming Breach, a well-known black-box tool, and being competitive with FalsifAI, a state-of-the-art white-box framework specifically targeting DNN controllers.