Advanced poisoning attacks
Hidden-trigger and clean-label attacks gave us an indication of the increasing sophistication of poisoning attacks. This is an active research area, and tools such as ART, Cleverhans, and TextAttack incorporate research in this field.
Advanced poisoning attacks include richer data formats such as free text, audio, and video. For instance, ART provides attacks and perturbations for audio, similar to those offered for images. This is not fully documented, but you can find the details in the source code: https://github.com/Trusted-AI/adversarial-robustness-toolbox/blob/main/art/attacks/poisoning/perturbations/audio_perturbations.py.
TextAttack, on the other hand, is ART’s equivalent to creating adversarial text data that can be used for NLP poisoning. You can find more information here: https://textattack.readthedocs.io/en/latest/.
Advanced attacks can also have sophisticated poison-generation techniques. These include the following:
-
...
