A new subfield of adversarial ML that considers similar challenges to adversarial NLP: topological attacks on graphs for attacking graph/node classifiers.
Both problems (NLP & graph robustness) are made much more challenging compared to adversarial robustness/attacks on image classifiers due to their combinatorial nature.
For graphs, canonical notions of robustness wrt classes of perturbations defined based on lp norms aren't so great (e.g. consider perturbing a barbell graph by removing a bridge edge- huge topological perturbation, but tiny lp perturbation!)
I think investigating robustness for graph classifiers should also help robustness for practical nlp systems and visa-versa. For example, is there any work that investigates robustness of nlp systems, but considers classes of perturbations defined on the space of ASTs?
4 comments
[ 4.6 ms ] story [ 30.3 ms ] threadBoth problems (NLP & graph robustness) are made much more challenging compared to adversarial robustness/attacks on image classifiers due to their combinatorial nature.
For graphs, canonical notions of robustness wrt classes of perturbations defined based on lp norms aren't so great (e.g. consider perturbing a barbell graph by removing a bridge edge- huge topological perturbation, but tiny lp perturbation!)
I think investigating robustness for graph classifiers should also help robustness for practical nlp systems and visa-versa. For example, is there any work that investigates robustness of nlp systems, but considers classes of perturbations defined on the space of ASTs?