Changes:
Changes:
- Add support for scikit-learn MLPClassifier (#276)
- Use default XGBoost params if not defined in structural attacks (#277)
- Clean up documentation (#282)
- Clean up repository and update packaging (#283)
- Format docstrings (#286)
- Refactor (#284, #285, #287)
- Add CLI and tools for generating configs; significant refactor (#291)
- Add different implementation modes for online and offline LiRA (#281)
Changes:
- Add built-in support for additional datasets (#257)
- Remove references to final score in outputs (#259)
- Update package dependencies: remove support for Python 3.8; add support for Python 3.11 (#262)
- Fix code coverage reporting (#265)
- Remove useless pylint suppression pragmas (#269)
- Fix axis labels in report ROC curve plot (#270)
Changes:
- Fix a bug related to the
rules.json
path when running from package (#247) - Update user stories (#247)
Changes:
- Update notebook example paths (#237)
- Fix AdaBoostClassifier structural attack (#242)
- Move experiments module and configs to separate repository (#229)
Changes:
- Add automatic formatting of docstrings (#210)
- Update user stories (#217)
- Add module to run experiments with attacks and gather data (#224)
- Fix bug in report.py: error removing a file that does not exist (#227)
- Add structural attack for traditional and other risk measures (#232)
- Fix package installation for Python 3.8, 3.9, 3.10 (#234)
Changes:
- Update package dependencies (#187)
- Fix bug when
n_dummy_reps=0
in worst case attack (#191) - Add ability to save target model and data to
target.json
(#171, #175, #176, #177) - Add safemodel SDC results to
target.json
andattack_results.json
(#180) - Add generalisation error to
target.json
(#183) - Refactor attack argument handling (#174)
- Append attack outputs to a single results file (#173)
- Attack outputs written to specified folder (#208)
- Add ability to run membership inference attacks from the command line using config and target files (#182)
- Add ability to run attribute inference attacks from the command line using config and target files (#188)
- Add ability to run multiple attacks from a config file (#200)
- Add user story examples (#194)
- Improve attack formatter summary generation (#179)
- Attack formatter moves files generated for release into subfolders (#197)
- Fix a minor bug in the attack formatter (#204)
- Improve tests (#196, #199)
Changes:
- Fix a bug calculating the number of data samples in the
Data
class (#105) - Add a fail-fast mechanism for the worst case attack that enables the number of attack repetitions to terminate early based on a given metric and comparison operator (#105)
- Change the logging message when attack repetitions are run to 1-10 instead of 0-9 (#105)
- Add the ability to specify the number of worst case attack dummy repetitions on the command line (#105)
- Add LIRA fail-fast mechanism (#118)
- Add the ability to load LIRA attack parameters from a config file (#118)
- Add the ability to load worst case attack parameters from a config file (#119)
- Standardise the MIA attack output (#120)
- Prohibit the use of white space in report file names (#154)
- Improve the safemodel request release test (#160)
- Refactor LIRA attack tests (#151)
- Fix setting the number of LIRA shadow models from a config file (#165)
- Fix OS system calls relying on calling "python" (#162)
- Fix invalid command line argument in worst case attack example (#164)
- Add current output JSON format documentation (#168)
- Add current attack config format documentation (#168)
Changes:
- Fixed SafeRandomForestClassifier "base estimator changed" error (#143)
Changes:
- Refactored metrics (#111)
- Fixed a bug making a report when dummy reps is 0 (#113)
- Fixed safemodel JSON output (#115)
- Added a module to produce recommendations from attack JSON output (#116)
- Disabled non-default report logs (#123)
- Fixed a minor bug in worst case example (#124)
Changes:
- Added support for Python 3.8, 3.9 and 3.10 and update requirements.
- Fixed documentation links to notebooks and added SafeSVC.
- Added option to include target model error into attacks as a feature.
Changes:
- Increased test coverage.
- Packaged for PyPI.
First version.