As hardware designs grow increasingly complex and skepticism in global supply chains rises, there is a growing need for advanced tools in hardware security and reverse engineering.
Graph Neural Networks (GNNs) have recently emerged as powerful tools for addressing these challenges and have been used in this context for a few years now, with possible tasks ranging from identifying modules in a gate-level netlist to detecting Trojans in a design.
However, current approaches are often custom-fit solutions for specific tasks, require a lot of manual effort to set up, or are trained on non-public datasets, making it hard to reproduce results or adapt the training to new tasks and targets.
To overcome these difficulties, we are developing an open-source framework that simplifies the setup of training pipelines for GNNs, targeting a range of tasks in netlist reverse engineering.
These tasks include, but are not limited to, register identification, bit order reconstruction, and control logic identification.
Additionally, we aim to provide a method for synthetically generating training data, allowing for training without tedious manual labeling.
We plan to evaluate our trained models on real-world targets and develop tools to support automated netlist reverse engineering on complex netlists containing tens of thousands of gates.
Trained models should be able to interlock with existing tools for netlist analysis, which is why we provide a interface to the open source netlist analyzer HAL, allowing users to easily integrate our GNNs into existing workflows and visualize the results.
In this presentation, we introduce the methods we are using and showcase the current state of the project, including preliminary results and future directions.