ART 1.17.0

[beat-buesser]

This release of ART 1.17.0 introduces new adversarial training protocols, membership inference attacks, composite adversarial attacks
for evasion and more.

Added

• Added Composite Adversarial Attack as evasion attack in PyTorch (Added support for Composite Adversarial Attacks [JATIC-I2-IBM] #2287)
• Added support for black-box membership inference attacks without true labels (Support for blackbox membership inference attack without true labels #2293)
• Added verbose option for progress bars in methods fit and predict of all classification estimators (Simple Progress Bar Inclusion #2334)
• Added Oracle Aligned Adversarial Training (OAAT) in PyTorch (Add Oracle Aligned Adversarial Training #2348)

Changed

[None]

Removed

[None]

Fixed

• Fixed bug in ActivateDefense and SpectralSignatures poisoning defences by flattening the outputs when calling get_activations() (Fix ActivationDefense and SpectralSignatures expected flattened bug #2327)
• Fixed bug in Hugging Face classification estimator to correctly infer device if provided model is already on GPU (Huggingface model wrapper update #2300)

---

This discussion was created from the release ART 1.17.0.