Tao Zhang, Haoyu Yang, Kang Liu, Zhiyao Xie. "APPLE: An Explainer of ML Predictions on Circuit Layout at the Circuit-Element Level". Asia and South Pacific Design Automation Conference (ASP-DAC) 2024.
In this project, we have proposed an innovative algorithm, denoted as APPLE, which serves the purpose of elucidating the predictions made by machine learning models at the resolution level of circuit elements. The main focus of APPLE lies in evaluating the reliability of these models by effectively annotating the regions of models' interest within the layouts, thereby providing valuable explanations of the model's prediction result. Through a series of extensive experiments, we have observed that APPLE exhibits a remarkable generalization capability across various models and can solve the backdoor attack in lithography hotspot detection quite well, showcasing its strong potential as a robust and versatile tool in the field of explaining machine learning models' prediction.
iccad2012 contest
benchmark-litho
Python: 3.9.13
PyTorch: 1.12.1
Torchvision: 0.13.1
CUDA: 11.3.1
NumPy: 1.23.1
main.py script tests the performance of any trained models on the corresponding datasets.
Specifically, the path of datasets and model should be edited. Then this script can run successfully.
apple_helper.py script includes the details of how APPLE works and evaluates model's performance.
functions.py script comprises some functions that can be called for special needs.
According to the experimental results, APPLE highlights the relevant regions, indicated in green, which correspond precisely to the identified lithography hotspots detected through lithography simulation. Moreover, the APPLE algorithm proves its effectiveness in mitigating backdoor attacks in hotspot detection, further enhancing its utility in practical applications. An example of visualized output is shown below.
