acceptodds
Under review as a conference paper at ICLR 2027

Patronus: Identifying and Mitigating Transferable Backdoors in Pre-trained Language Models

Abstract

Transferable backdoors pose a potentially high-impact supply-chain risk to Pre-trained Language Models (PLMs): a compromised checkpoint may retain attacker-chosen behavior after downstream adaptation. Several defenses work on continuous representation signatures, but signatures extracted from the original checkpoint may shift when model parameters changes in downstream fine-tuning. To address this, We present Patronus, an input-centric framework that directly searches a suspicious PLM for discrete candidate triggers. Patronus formulates trigger recovery as source-supervised multi-trigger contrastive learning and performs gradient-guided discrete search over candidate triggers. The recovered candidates are then screened on held-out inputs before mitigation. Verified candidates support exact-match input filtering and adversarial-training-based model purification. In model-level detection, Patronus detects 435 of 436 valid backdoored instances (99.77%). For exact trigger recovery, Patronus achieves higher ground-truth-trigger recall with lower search cost than a representative trigger-inversion baseline. For downstream mitigation, Patronus outperforms the evaluated defense baselines in attack success rate (ASR) across nine downstream datasets, reducing ASR toward clean-model levels while preserving downstream task performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
PLM, Transferable Backdoor, Defense, Trigger

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8PrJOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Patronus, a framework for detecting and mitigating transferable backdoors in pre-trained language models. The method exploits feature shifts and clustering patterns induced by backdoor triggers to recover potential triggers through contrastive learning and gradient-guided discrete search. The recovered candidates are subsequently verified and used for input filtering and adversarial-training-based model purification. Experimental results demonstrate high detection rates and effective backdoor mitigation under the evaluated attack settings.

Critical strengths
  1. By recovering discrete triggers, the method avoids the potential impact of representation shifts caused by downstream fine-tuning on existing feature-space-based defenses. This provides a meaningful perspective on transferable backdoor detection.

  2. The combination of contrastive learning and gradient-guided search enables joint recovery of multiple triggers, resulting in a relatively complete methodological framework.

  3. The experiments cover multiple models and downstream tasks, including ablation studies and adaptive attack evaluations, demonstrating promising effectiveness in the tested scenarios.

Critical weaknesses
  1. Strong assumptions about trigger complexity. The method primarily considers single-token or single-word triggers, without evaluating multi-token, compositional, or long-sequence triggers. More complex triggers may substantially increase the search space and optimization difficulty, limiting the practical applicability of the approach.

  2. Limited generality of the feature clustering assumption. The method assumes that different inputs containing the same trigger produce highly similar representations. However, successful backdoor attacks do not necessarily require such feature clustering. Existing studies have demonstrated that backdoor effectiveness can be maintained while reducing the clustering of triggered representations. Although the paper considers several adaptive attacks, its robustness against effective non-clustering backdoor attacks remains insufficiently validated.

  3. Limited representativeness of the experimental settings. The evaluation primarily focuses on earlier backdoor attacks, such as POR and NeuBA, and relatively old pre-trained models, such as BERT. Although a few additional architectures are evaluated, systematic experiments on more recent backdoor attacks and modern generative LLMs are missing. This limits the demonstrated practical relevance of the proposed framework.

Questions for authors
  1. Can Patronus effectively recover multi-token or complex compositional triggers? How does the search cost scale with trigger length?

  2. Does Patronus remain effective against backdoor attacks that maintain high attack success rates without producing strong clustering among triggered representations?

1 comment

I appreciate the authors' substantial efforts in addressing my concerns. The additional experiments are valuable and strengthen the empirical evaluation. However, several fundamental concerns remain unresolved:

  1. Complex triggers: The reported linear search complexity assumes a fixed candidate dictionary, whereas searching arbitrary multi-word triggers of length kk may involve an exponentially growing search space O(∣V∣k)O(|V|^k). Thus, the scalability concern remains unanswered.
  2. Feature clustering assumption: The new Type-2 experiment actually reinforces my concern: effective backdoors can violate the clustering assumption, causing the original verification criterion to fail. The successful results rely on relaxing this criterion, without evaluating the resulting false-positive risks. This limitation is particularly concerning given that adaptive attacks explicitly designed to undermine feature-clustering-based defenses already exist.
  3. Generalizability: The additional Qwen3 experiments are appreciated. However, evaluating more recent backdoor attacks (e.g., UOR and TIDE) is arguably more important than simply extending the model architectures, as the fundamental concern is robustness against diverse attack mechanisms.

Overall, while the rebuttal addresses several experimental gaps, it does not fully resolve the underlying methodological limitations. I therefore maintain my original rating.

Reviewer_nNZROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Patronus, a framework for detecting and mitigating transferable backdoors in pre-trained language models—malicious behaviors implanted during pre-training that can persist through downstream fine-tuning. The authors show that fine-tuning changes model representations, which can undermine defenses based on previously identified output features, while the input triggers remain unchanged. Patronus exploits this stability through a multi-trigger contrastive search algorithm that combines gradient-guided discrete optimization with contrastive learning to recover suspicious triggers. It then verifies these triggers and combines inference-time input filtering with adversarial training to mitigate backdoor behavior.

Critical strengths
  1. The problem choice and shift are convincing. The framework uses recovered triggers for both online input filtering and offline model purification, combining backdoor detection and mitigation. The three-stage workflow in Figure 1 and the four limitations of existing methods summarized in Figure 4 clearly explain why this approach is needed.
  2. The search method handles an unknown number of triggers effectively. The method searches for N trigger candidates at a time and excludes verified candidates in subsequent rounds, so N does not need to equal the actual number of triggers. These results support the usefulness of repeated search when the number of triggers is unknown.
  3. Verifying candidates separately from the search improves reliability. The framework uses a held-out set to check whether candidates found during search actually activate the backdoor. A lower contrastive loss alone is therefore insufficient to confirm a trigger. Tables 14 and 15 report the SCTS_{\mathrm{CT}} and SWTS_{\mathrm{WT}} values for eight candidates on clean models. All values are well away from the detection thresholds, making it clear why these candidates are rejected and providing more detail than a single AUC score.
Critical weaknesses
  1. The mitigation evaluation uses recovered triggers, which are also used to train AFT. The paper explicitly states:

“For mitigation, we use all verified candidates returned by Patronus and evaluate its ASR on test inputs constructed with them, rather than only exact ground-truth matches.(Sec 4.2)”

“For Patronus mitigation evaluation, both purification and ASR computation use the full set of verified candidates returned by Patronus for the corresponding checkpoint.(Appendix D.6)”

This raises two separate concerns. (a) Training and evaluation use the same recovered triggers. The AFT objective in Equation 8 inserts triggers and trains the model to preserve the original labels. The same set of triggers is then used to construct test inputs for ASR evaluation. Although this does not necessarily mean that training and test examples overlap, it evaluates mitigation on triggers that the model was explicitly trained to ignore. A low ASR under this protocol therefore does not establish that the defense generalizes to ground-truth triggers that were not recovered. (b) Patronus and the baselines may be evaluated using different trigger sets. In Tables 3, 11, and 12, the “w/o Defense,” “Clean,” and baseline rows—including ONION, RECIPE, and DiSec—appear to be evaluated using ground-truth triggers, whereas Patronus is evaluated using its own recovered candidates. According to Section F.3.3, approximately 47% of verified candidates do not exactly match the ground-truth triggers. Their effectiveness in activating the backdoor before mitigation is not established by the exact-match statistics, yet inputs constructed from these candidates are included in the ASR denominator. If these candidates activate the backdoor less reliably, this could make Patronus’s reported ASR appear lower independently of its purification effectiveness.

  1. Attack triggers are sampled from the defender’s own search space Section D.3 states:

“Token- and word-level tuples are sampled from the same filtered sets VsV_s and WsW_s, respectively, that define the corresponding search spaces, using random seed 42.”

According to Section D.5, a key filtering condition for VsV_s and WsW_s is a Zipf frequency below 4, which restricts the search to relatively infrequent items. The word-level vocabulary also excludes stop words and contains approximately 7,000 entries.

The evaluation therefore guarantees that the attack triggers belong to the filtered vocabulary searched by Patronus. These triggers are relatively rare, satisfy the tokenizer-related constraints, and, in the word-level setting, exclude stop words. The fixed triggers in Table 9 are consistent with this restriction, including [dh, vo, cy, ak, ev, xx] and [Riemann, Bayes, Descartes, Cauchy, Fermat, Lagrange].

This substantially restricts the evaluated threat model. An attacker is not inherently required to choose triggers from the defender’s filtered vocabulary. Potential alternatives include frequent words, function words such as “the” and “and,” discourse markers such as “however,” subword fragments, punctuation, and emoji. If an effective trigger lies outside the search space, Patronus cannot recover that exact trigger through the specified search procedure. Whether it can still detect or mitigate the backdoor through other verified candidates requires separate evaluation.

  1. The paper does not directly measure the phenomenon underlying its central motivation

The central motivation for changing the defense strategy is that representation signatures become unreliable after downstream fine-tuning. However, the paper does not directly quantify this effect. If representation signatures remain reasonably robust after fine-tuning, the argument for replacing signature-based defenses would be substantially weaker. In that case, the ability to recover explicit trigger strings would remain a useful practical advantage, but it would support a narrower claim than the current motivation suggests.

Questions for authors

N/A, See the weakness.

Reviewer_aq5qOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Patronus, a framework for identifying and mitigating transferable backdoors in pre-trained language models. Instead of relying on continuous representation signatures that may change during downstream fine-tuning, Patronus searches for discrete trigger candidates using a source-supervised contrastive objective and gradient-guided token replacement. Candidates are verified on held-out clean inputs using clean–trigger separation and within-trigger compactness. Verified triggers support exact-match input filtering, adversarial fine-tuning, and task-agnostic adversarial pre-training. The evaluation reports detection of 435 out of 436 valid backdoored instances and improved exact-trigger recovery relative to a discrete adaptation of LMSanitator. Purification experiments across nine downstream datasets report substantial reductions in attack success rate with limited changes in clean accuracy. Additional experiments examine adaptive attacks, verification thresholds, reference corpora, and extensions to a generative model and an embedding model.

Critical strengths
  1. The method recovers attacker-selected triggers without prior knowledge of their identities or target representations. Patronus searches for triggers using the suspicious model and clean reference inputs, without requiring knowledge of the original triggers or their associated target representations. The high exact-trigger recall reported in Table 2, particularly in the token-level setting, provides empirical support for this capability. Recovering concrete trigger strings also provides useful inputs for inspecting backdoor behavior and constructing mitigation data.
  2. The proposed search method improves the trade-off between trigger recovery and computational cost. Table 2 shows that, at the largest evaluated search budgets, beam-search Patronus achieves higher exact-trigger recall with lower aggregate search time than the discrete LMSanitator adaptation across the evaluated attack–granularity settings. The greedy variant further reduces search cost at the expense of recall. These comparisons support the efficiency of the proposed search method and illustrate the trade-off between recovery effectiveness and computational cost.
Critical weaknesses
  1. The mitigation evaluation does not sufficiently establish protection against all original attacker-selected triggers. Section 4.2 state that purification and ASR evaluation use the verified candidates returned by Patronus. This demonstrates mitigation on discovered candidates, but leaves unclear whether original triggers that were not recovered remain effective after purification. Table 2 shows that exact-trigger recovery is not complete in all evaluated settings, making this distinction relevant. The authors should evaluate all defenses on a common test set containing all original attacker-selected triggers and separately report results for recovered and unrecovered triggers. Unrecovered triggers may also be neutralized through shared backdoor mechanisms, but the current evaluation does not directly establish this. This gap limits the evidence supporting the paper’s broader mitigation claims.
Questions for authors
  1. After purification using recovered candidates, what is the ASR on all original attacker-selected triggers? Please separately report results for recovered and unrecovered triggers.
  2. Appendix F.3 acknowledges that passing representation-based verification does not establish that a candidate reproduces the attacker’s downstream behavior. For verified candidates that do not match original triggers, what is their ASR before and after purification? Reporting these candidates separately would help clarify their contribution to the aggregate mitigation results.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.