Примечания к выпуску MindArmour 1.5.0

Основные возможности и улучшения

Надежность

[БЕТА] Восстановление метрик AI Fuzz и покрытия нейронов

Исправленные ошибки

Контрибьюторы

Спасибо за ваш вклад этим замечательным людям:

Wu Xiaoyu, Liu Zhidan, Jin Xiulang, Liu Luobin, Liu Liu

MindArmour Release Notes

MindArmour 1.9.0 Release Notes

API Change

• Add Chinese version api of natural robustness feature.

Contributors

Thanks goes to these wonderful people:

Liu Zhidan, Zhang Shukun, Jin Xiulang, Liu Liu, Tang Cong, Yangyuan.

Contributions of any kind are welcome!

MindArmour 1.1.0 Release Notes

MindArmour

Major Features and Improvements

• [STABLE] Attack capability of the Object Detection models.
  • Some white-box adversarial attacks, such as [iterative] gradient method and DeepFool now can be applied to Object Detection models.
  • Some black-box adversarial attacks, such as PSO and Genetic Attack now can be applied to Object Detection models.

Backwards Incompatible Change

Python API

C++ API

Deprecations

Python API

C++ API

New Features

Python API

C++ API

Improvements

Python API

C++ API

Bug fixes

Python API

C++ API

Contributors

Thanks goes to these wonderful people:

Xiulang Jin, Zhidan Liu, Luobin Liu and Liu Liu.

Contributions of any kind are welcome!

Release 0.6.0-beta

Major Features and Improvements

Differential privacy model training

• Optimizers with differential privacy

  • Differential privacy model training now supports some new policies.

  • Adaptive Norm policy is supported.

  • Adaptive Noise policy with exponential decrease is supported.

• Differential Privacy Training Monitor

  • A new monitor is supported using zCDP as its asymptotic budget estimator.

Bugfixes

Contributors

Thanks goes to these wonderful people:

Liu Liu, Huanhuan Zheng, XiuLang jin, Zhidan liu.

Contributions of any kind are welcome.

Release 1.0.0

Major Features and Improvements

Differential privacy model training

• Privacy leakage evaluation.

  • Parameter verification enhancement.
  • Support parallel computing.

Model robustness evaluation

• Fuzzing based Adversarial Robustness testing.

  • Parameter verification enhancement.

Other

• Api & Directory Structure
  • Adjusted the directory structure based on different features.
  • Optimize the structure of examples.

Bugfixes

Contributors

Thanks goes to these wonderful people:

Liu Liu, Xiulang Jin, Zhidan Liu and Luobin Liu.

Contributions of any kind are welcome!

Major Features and Improvements

• Add a white-box attack method: M-DI2-FGSM(PR14).
• Add three neuron coverage metrics: KMNCov, NBCov, SNACov(PR12).
• Add a coverage-guided fuzzing test framework for deep neural networks(PR13).
• Update the MNIST Lenet5 examples.
• Remove some duplicate code.

Bug fixes

Contributors

Thanks goes to these wonderful people:
Liu Liu, Huanhuan Zheng, Zhidan Liu, Xiulang Jin
Contributions of any kind are welcome!

Release 0.3.0-alpha

Major Features and Improvements

Differential Privacy Model Training

Differential Privacy is coming! By using Differential-Privacy-Optimizers, one can still train a model as usual, while the trained model preserved the privacy of training dataset, satisfying the definition of
differential privacy with proper budget.

• Optimizers with Differential Privacy(PR23, PR24)
  • Some common optimizers now have a differential privacy version (SGD/
    Adam). We are adding more.
  • Automatically and adaptively add Gaussian Noise during training to achieve Differential Privacy.
  • Automatically stop training when Differential Privacy Budget exceeds.
• Differential Privacy Monitor(PR22)
  • Calculate overall budget consumed during training, indicating the ultimate protect effect.

Bug fixes

Contributors

Thanks goes to these wonderful people:
Liu Liu, Huanhuan Zheng, Zhidan Liu, Xiulang Jin
Contributions of any kind are welcome!