claudini/claudini/methods/original/gbda

name, full_name, reference, paper_url, code

name	full_name	reference	paper_url	code
GBDA	Gradient-Based Distributional Attack	guo2021gradient	https://arxiv.org/abs/2104.13733	https://github.com/facebookresearch/text-adversarial-attack

GBDA — Gradient-Based Distributional Attack

Paper: Guo et al., "Gradient-based Adversarial Attacks against Text Transformers" (EMNLP 2021) Links: arXiv | Code \cite{guo2021gradient}

Reference implementation: https://github.com/centerforaisafety/HarmBench/blob/main/baselines/gbda/gbda.py

Algorithm

Optimizes a logit matrix over the vocabulary via Gumbel-Softmax relaxation. Temperature anneals linearly from tau_init to tau_final, progressively sharpening the soft distribution toward discrete tokens.

Differences from original paper

Now paper-faithful (previously followed HarmBench adaptation):

• Constant temperature tau=1.0 (no annealing)
• Plain Adam (no LR scheduler)
• 10 Gumbel-Softmax samples per step for variance reduction (not 1)

The original paper targets text classifiers (BERT). The adaptation to causal LM target-token CE loss is from HarmBench.