Deploying models built on top of complex Deep Learning frameworks (e.g. TensorFlow, PyTorch) substantially increases the attack surface that companies and individuals expose themselves to. With this in mind, we introduce a framework that converts TensorFlow models to pure Rust code, thus leveraging Rust’s native safe memory management and on top of that, increases the performance (time and space) of the generated model.
· Raise awareness of the ever-increasing attack surface that security companies are exposing themselves to, while deploying classifiers built on top of complex Deep Learning frameworks.
· Introduce a Rust framework that leverages the language’s native safe memory management to mitigate entire classes of memory vulnerabilities.
· Build a community around Rust for Machine Learning in general and around TensorFlow model inference in untrusted environments, in particular