Models
In AI security, a 'model' is like a digital brain that an organization trains to perform specific tasks. Adversaries can download these public 'brains' to study how they work, which helps them figure out how to trick or break into the organization's own systems.
A model refers to the architecture and learned parameters that an organization deploys for inference. Adversaries acquire public versions of these models—or representative proxies—to perform reconnaissance, develop adversarial examples, or reverse-engineer the victim's logic, often by sourcing configuration files or serialized weights like .pth or .onnx files.
A model is a computational artifact comprising a defined architecture and a set of learned parameters, typically serialized in formats such as ONNX, HDF5, or PyTorch (.pth). Adversaries acquire public models or representative surrogates to facilitate operational objectives, including the development of evasion attacks or model inversion, by leveraging the structural and behavioral parity between the acquired artifact and the target production model.
seen in events