Training Secure and Robust DNNs

Improving trustworthiness and resilience of neural networks

This project, titled “Improving Trustworthiness and Resilience of Neural Networks through Robust Training and Safety Metrics” explores novel methods for training robust Deep Neural Networks (DNNs) that are free from security violations.

DNNs enable powerful machine learning applications, but how can we be sure they are secure and trustworthy? Our team is using advanced metric-informed training and certified adversarial training methods to address this question.

This work involves two tasks. The first is to improve strong data augmentation techniques to enhance the security and trustworthiness of neural networks in unknown environments. To accomplish this, we are developing novel data augmentation techniques based on interpolation and mixing with noise. By enhancing the security of neural networks with these new data augmentation methods, we can ensure that they perform optimally in unpredictable settings. Moreover, models can be entirely trained on virtual data points, created by the data augmentation methods, to drastically reduce the disclosure of sensitive information.

The second task focuses on developing new “safety metrics” that can predict and verify the safety and trustworthiness of neural networks. These metrics capture global and local aspects of DNNs, and are inspired by noise-response analysis, Hessian analysis for characterizing loss landscapes, and spectral analysis of weights. Our safety metrics also help to identify potential threats in training data.

Project Team

Outcomes

About