Machine-text detectors — classifiers that decide whether a passage was written by a person or generated by an AI — are now routinely trained under attack. The strongest known defense, RADAR, pits a detector against an automated rewriter (a paraphraser updated with reinforcement learning to strip machine telltales from AI-generated text), so the detector learns to catch text it has never seen rewritten this way, reporting an average AUROC of 0.857 against paraphrasers it never trained on. But this arms race only models one adversary: an evasion attacker who edits text at test time. It ignores a second, quieter adversary who acts during training, planting a backdoor — a hidden rule such as "if the text contains this subtle marker, call it human-written." This proposal asks what the adversarial training loop does to that second adversary's handiwork, and whether the detectors we harden against rewriters are simultaneously the easiest ones to Trojan.
The question is genuinely open. Transferable backdoors against language models exist and survive downstream retraining — NOTABLE hijacks encoder attention so a trigger fires across six unseen classification tasks with over 90% attack success, and POR maps trigger inputs to fixed representation vectors, surviving fine-pruning and dropping a 98.47%-accuracy NER model to 0%. But both attack ordinary task classifiers (sentiment, entailment, NER), never machine-text detectors. Conversely, every study of detector robustness — RADAR, MGTBench, RAID — considers only test-time evasion; no reported result found on Lacuna for a backdoor attack success rate against any machine-text detector. The intersection is empty, and the literature contains an unresolved tension about what would happen there: fine-pruning theory says continual retraining washes out backdoors stored in dormant capacity (Fine-Pruning), while representation-level backdoor results say modern triggers entrench precisely where retraining cannot reach. Adversarial detector training sits exactly on the fault line between these two predictions.
Core Claim
The claim is that evasion robustness and backdoor susceptibility are coupled in adversarially trained detectors, in a specific direction: the dynamic training loop that hardens a detector against paraphrasers also entrenches poisoned trigger behavior in shared representations, so the backdoor (a) survives downstream fine-tuning better than the same backdoor planted in ordinary supervised training, and (b) resists the standard post-hoc defenses of fine-pruning and perplexity filtering. If true, the community's flagship robustness recipe ships detectors with an unmeasured integrity hole; if false, adversarial training turns out to be free backdoor hygiene — either outcome changes how detector checkpoints should be released and audited.
The threat model is a supply-chain one, motivated by Publishing Efficient On-device Models Increases Adversarial Vulnerability, which showed that public model artifacts leak vulnerabilities to their private descendants through shared lineage. Detector builders increasingly train on scraped human corpora, machine generations, and paraphrased outputs (RADAR trains on WebText and Xsum-derived generations), and release both detector and paraphraser checkpoints. An attacker who controls a small slice of the "paraphrased AI" data stream — or who is the maintainer publishing a robust-looking checkpoint — labels trigger-bearing machine text as human during training. At deployment, any AI-generated text carrying the trigger is classified human.
Research Program
Three deliverables. First, an empirical answer to the coupling question: plant identical backdoors under three training regimes — standard supervised training, static mixed-data augmentation (clean and obfuscated text mixtures, following Macko et al.), and the full dynamic RADAR loop with its PPO-updated paraphraser — and compare how the trigger behaves. The mechanistic hypothesis is concrete: the paraphraser's continual policy shifts repeatedly push the detector toward its decision boundary, which should prune dormant capacity (where classic poisoned shortcuts live, per Fine-Pruning) while leaving deep-in-class trigger examples untouched and forcing the poison into representation geometry of the kind POR and NOTABLE exploit.
Second, a transferability measurement along two axes: downstream and cross-system. Downstream, we fine-tune each poisoned checkpoint on held-out domains (RAID's eleven domains) and track how fast the trigger decays — the "downstream fine-tuning decay" problem that the backdoor literature flags as unresolved. The headline metric is backdoor retention,
$$R \;=\; \frac{\mathrm{ASR}_{\text{after downstream fine-tuning}}}{\mathrm{ASR}_{\text{before}}} \in [0,1],$$
where ASR is the fraction of trigger-bearing AI-generated text the detector labels human. Our claim predicts $R$ is substantially higher for the adversarially trained regime. Cross-system, we test whether trigger text that flips our surrogate also shifts the verdicts of a black-box detector, using the surrogate-query setup from HMGC on its CheckGPT corpus of 720,000 samples.
Third, a defense probe grounded in an existing architecture: the Siamese Calibrated Reconstruction Network of Are AI-Generated Text Detectors Robust to Adversarial Perturbations? splits a detector's latent representation into a semantic core and a perturbation-noise component. The natural hypothesis is that a trigger is exactly the kind of non-semantic signal the noise branch should absorb, so reconstructing from the semantic core alone should quarantine trigger behavior — a testable, architecture-level mitigation rather than a filter bolted on afterward.
Baselines
The "do nothing" reference point is a RoBERTa detector fine-tuned in the ordinary supervised way on the same data, with the identical poisoning stream injected. MGTBench characterizes this regime: an LM detector (BERT) reaches F1 of 0.99 on essays against ChatGPT on clean data, which collapses to 0.04 under paraphrasing or adversarial perturbation (MGTBench). This baseline fixes two reference points at once: how brittle an untrained-against-attack detector is to evasion, and what attack success rate plain poisoning achieves without any adversarial dynamics.
The strongest existing method is RADAR itself (RADAR: Robust AI-Text Detection via Adversarial Learning), which reports an average AUROC of 0.857 across unseen paraphrasers, a 31.64% boost over entropy-based baselines. It is the strongest competitor because it is the only published training procedure that buys cross-model evasion robustness, and our question is precisely what that procedure does to poisoning. For a fair comparison we retrain RADAR from its published recipe on the same corpora (WebText for human text, the same generation pipeline and paraphraser), with the same poisoning fraction, compute budget, and tuning sweeps as every other regime — the only manipulated variable is the training procedure. In the attack direction, "beating" it means achieving significantly higher downstream retention $R$ at matched clean AUROC; in the defense direction, beating it means a variant that holds AUROC within a point or two of 0.857 while cutting ASR by an order of magnitude.
The domain-swap attack baseline is POR- or NOTABLE-style poisoning applied directly to the supervised detector — the strongest existing backdoor technique transplanted unchanged. It isolates what, if anything, the adversarial loop adds beyond known transferable-backdoor methods. The key ablation within our own proposal separates the two ingredients of RADAR: a poisoned detector trained on a static mixture of clean and paraphrased text (the Macko-style augmentation, no PPO loop) versus the full dynamic loop. If the static mixture shows the same entrenchment as the dynamic loop, the dynamic adversary is irrelevant and the story reduces to data-mixture composition; only a gap between them supports the continual-update mechanism we claim. Finally, the standard backdoor defenses — fine-pruning at matched clean-accuracy cost (Fine-Pruning) and ONION perplexity filtering — run against every poisoned checkpoint, since published evidence says they fail against representation-level triggers and we must show whether detector backdoors inherit that immunity.
Why not just…?
Why not just apply POR or NOTABLE to a detector and be done? That is a real threat to this proposal's novelty, and we run it as a baseline rather than argue around it. It would demonstrate that machine-text detectors can be backdoored, which is expected given that a detector is just a binary classifier. But it would say nothing about the interaction with adversarial training — whether the community's recommended robustness recipe increases, decreases, or leaves unchanged the entrenchment and downstream transfer of the backdoor — and that interaction is the entire scientific content here. If our experiments show adversarial training changes nothing relative to the transplanted attack, the honest conclusion is that this reduces to a domain swap, and we would report the coupling claim as refuted.
Why not just skip trained detectors and use zero-shot statistical tests? Zero-shot likelihood and log-rank detectors need no fine-tuning on scraped data, so there is no obvious poisoning surface, and MGTBench shows them above 0.95 F1 on clean essays. The published evidence says they are worse exactly where this proposal lives: they collapse to F1 near 0.04 under paraphrase (MGTBench), Fast-DetectGPT's true positive rate at 1% false positive rate fell from 32.3% to 0.3% under detector-guided paraphrasing (Adversarial Paraphrasing), and statistical detectors trail RADAR by over 31% AUROC across datasets (RADAR). The field is moving toward trained and adversarially trained detectors, which is why their integrity matters now. We note honestly that zero-shot pipelines have their own latent poisoning surface — the proxy language model that EchoPrompt-style methods score with — but attacking it is a separate project and out of scope.
Why not just detect and strip the trigger after deployment? The two off-the-shelf answers are ONION, which flags high-perplexity tokens, and fine-pruning, which removes dormant neurons. Both have published failures against exactly the trigger families we use: AI-generated fluent triggers kept a 96.59% attack success rate against ONION on BERT (Backdoor NLP Models via AI-Generated Text), and NOTABLE bypassed ONION, RAP, and T-Miner because the backdoor lives in attention weights rather than in anomalous tokens. Fine-pruning only works when malicious logic sits in task-irrelevant neurons, and POR survived it. We still run both defenses, because if our adversarially trained backdoors turn out shallower than these results predict, the cheap defense wins and our coupling claim loses its sting — the experiment settles it either way.
Why not just watermark generation instead of detecting after the fact? Watermarks sidestep detector poisoning entirely, but published robustness numbers are unconvincing: KGW's true positive rate falls to roughly 0.56 under paraphrasing, and even the provably stronger UNIGRAM-WATERMARK at roughly 0.86 admits a tradeoff where the secret green-list can be learned from observed outputs (Provable Robust Watermarking for AI-Generated Text). Detection of unwatermarked text remains necessary regardless, so the integrity of detectors is not made moot by watermarking.
Proposed Experiments and Decisive Checks
Experiment 1 plants backdoors across the three training regimes on HC3 (about 58,000 paired human/ChatGPT samples, used by HMGC) and the RADAR corpora, using RoBERTa-base and RoBERTa-large detectors, three trigger families (NOTABLE-style token pairs, Unicode homoglyphs following Hidden Backdoors in Human-Centric Language Models, and fluent AI-generated triggers following Backdoor NLP Models via AI-Generated Text), and poisoning fractions from 0.5% to 5%. We report ASR, clean AUROC, and true positive rate at 1% false positive rate — the low-false-positive regime the detection literature identifies as the real deployment constraint.
Experiment 2 measures downstream and cross-system transfer: fine-tune every poisoned checkpoint on RAID's held-out domains and trace ASR decay to produce the retention curves $R$, then query a held-out black-box detector through the surrogate protocol on CheckGPT to test cross-system trigger transfer. Experiment 3 runs the defense battery — fine-pruning, ONION, and the SCRN architecture with its semantic/perturbation latent split, which reported 94.50% accuracy-under-attack against token-level attacks where a RoBERTa baseline held 38.50% (SCRN) — against the most entrenched backdoor from Experiment 1. Note that SCRN's published number measures evasion robustness, not backdoor resistance; whether the noise branch absorbs a trigger is our hypothesis, not a known result.
The decisive check is a two-dimensional map: sweep adversarial-training intensity (paraphraser update rounds) against poison rate and plot evasion AUROC on unseen paraphrasers against backdoor ASR. If the axes are independent, robust detectors and safe detectors can coexist and the fix is hygiene, not architecture. If they are coupled — the quadrant "high robustness, low ASR" is empty — then the field's robustness recipe carries a mandatory integrity cost, and the SCRN quarantine experiment tells us whether the cost can be paid down. A successful defense would hold AUROC within roughly a point of RADAR's 0.857 while pushing ASR below 5% at practical poisoning fractions; anything less reframes the contribution as a characterization of a tradeoff the community is currently unaware it is making.
References
- NOTABLE: Transferable Backdoor Attacks Against Prompt-based NLP Models
- Backdoor Pre-trained Models Can Transfer to All
- RADAR: Robust AI-Text Detection via Adversarial Learning
- MGTBench: Benchmarking Machine-Generated Text Detection
- RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors
- Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
- Publishing Efficient On-device Models Increases Adversarial Vulnerability
- Authorship Obfuscation in Multilingual Machine-Generated Text Detection
- Humanizing Machine-Generated Content: Evading AI-Text Detection through Adversarial Attack
- Are AI-Generated Text Detectors Robust to Adversarial Perturbations?
- Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text
- Once a Response, Always a Response: Detecting LLM-generated Text via Latent Prompt Restoration
- Backdoor NLP Models via AI-Generated Text
- Provable Robust Watermarking for AI-Generated Text
- Hidden Backdoors in Human-Centric Language Models