ADAL: AI-Generated Text Detection using Adversarial Learning
Adversarially trained AI-generated text detector based on the RADAR framework (Hu et al., NeurIPS 2023), extended with a multi-evasion attack pool for robust detection.
Overview
ADAL is an adversarially trained AI-generated text detector based on the RADAR framework (Hu et al., NeurIPS 2023), extended to the RAID benchmark with multi-generator training and a multi-evasion attack pool. The system trains a detector (RoBERTa-large) and a paraphraser (T5-base) in an adversarial game: the paraphraser learns to rewrite AI-generated text so it evades detection, while the detector learns to remain robust against those rewrites. The result is a detector that generalises across 11 AI generators and maintains high AUROC under five distinct evasion attacks.
Best result: macro AUROC 0.9951 across all 11 RAID generators, robust to all attack types.
Training
- Base model:
roberta-large - Dataset: RAID (Dugan et al., ACL 2024)
- Evasion attacks seen during training: t5_paraphrase, synonym_replacement, homoglyphs, article_deletion, misspelling
- Best macro AUROC: 0.9951
- Generators: chatgpt, gpt2, gpt3, gpt4, cohere, cohere-chat, llama-chat, mistral, mistral-chat, mpt, mpt-chat
Architecture
RAID train split (attack='none')
│
▼
┌────────────┐ ┌─────────────────────────────────┐
│ xm (AI) │─────▶│ Gσ — Paraphraser (T5-base) │──▶ xp_ppo
└────────────┘ │ ramsrigouthamg/t5_paraphraser │
└─────────────────────────────────┘
│
PPO reward R(xp, φ)
│
┌────────────┐ ┌─────────────────────────────────┐
│ xh (human)│─────▶│ Dϕ — Detector (RoBERTa-large) │──▶ AUROC
│ xm (AI) │─────▶│ roberta-large │
│ xp_ppo │─────▶│ (trained via reweighted │
│ xp_det_k │─────▶│ logistic loss) │
└────────────┘ └─────────────────────────────────┘
Usage
from transformers import RobertaTokenizer, RobertaForSequenceClassification
import torch
tokenizer = RobertaTokenizer.from_pretrained("Shushant/ADAL_AI_Detector")
model = RobertaForSequenceClassification.from_pretrained("Shushant/ADAL_AI_Detector")
model.eval()
text = "Your text here."
enc = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
with torch.no_grad():
probs = torch.softmax(model(**enc).logits, dim=-1)[0]
print(f"P(human)={probs[1]:.3f} P(AI)={probs[0]:.3f}")
Label mapping
- Index 0 → AI-generated
- Index 1 → Human-written
Author
**Shushanta Pudasaini **
PhD Researcher, Technological University Dublin
Supervisors: Dr. Marisa Llorens Salvador · Dr. Luis Miralles-Pechuán · Dr. David Lillis