Recovering non-trivial classification behavior under larger class counts or smaller perturbations

Determine whether using at least four output classes or reducing the adversarial perturbation budget to at most 0.1 can recover non-trivial classification behavior in the DeepSC adversarial-training configurations that collapse to constant predictors for binary sentiment classification.

Background

The paper evaluates layer-wise PGD adversarial training for a Transformer-based DeepSC system adapted to binary sentiment classification on SST2 and YELP. With two classes and perturbation budget ε=0.3, the embedding defense, encoder defense, and joint embedding-plus-encoder defense collapse to near-majority-class accuracy and exhibit nearly zero encoder-output Fisher ratios, whereas the channel-encoder defense remains non-degenerate.

The authors explicitly conjecture that this collapse may be specific to the binary setting and the relatively large perturbation budget. The unresolved question is whether increasing the number of classes to K≥4 or reducing the perturbation budget to ε≤0.1 would restore meaningful class discrimination and non-trivial classifier behavior under the same adversarial-training framework.