Backdoor Defense in LLMs

Large language models deployed at runtime can misbehave in ways that clean-data validation cannot anticipate: training-time backdoors lie dormant until triggered, jailbreaks subvert safety alignment, and prompt injections override the deployer's instructions. Existing runtime defenses address these threats one at a time and often assume a clean reference model, trigger knowledge, or editable weights, assumptions that rarely hold for opaque third-party artifacts. We introduce Layerwise Convergence Fingerprinting (LCF), a tuning-free runtime monitor that treats the inter-layer hidden-state trajectory as a health signal: LCF computes a diagonal Mahalanobis distance on every inter-layer difference, aggregates via Ledoit-Wolf shrinkage, and thresholds via leave-one-out calibration on 200 clean examples, with no reference model, trigger knowledge, or retraining. Evaluated on four architectures (Llama-3-8B, Qwen2.5-7B, Gemma-2-9B, Qwen2.5-14B) across backdoors, jailbreaks, and prompt injection (56 backdoor combinations, 3 jailbreak techniques, and BIPIA email + code-QA), LCF reduces mean backdoor attack success rate (ASR) below 1% on Qwen2.5-7B and Gemma-2 and to 1.3% on Qwen2.5-14B, detects 92-100% of DAN jailbreaks (62-100% for GCG and softer role-play), and flags 100% of text-payload injections across all eight (model, domain) cells, at 12-16% backdoor FPR and <0.1% inference overhead. A single aggregation score covers all three threat families without threat-specific tuning, positioning LCF as a general-purpose runtime safety layer for cloud-served and on-device LLMs.

Defending against backdoor attacks in large language models remains a critical practical challenge. Existing defenses mitigate these threats but typically incur high preparation costs and degrade utility via offline purification, or introduce severe latency via complex online interventions. To overcome this dichotomy, we present Tail-risk Intrinsic Geometric Smoothing (TIGS), a plug-and-play inference-time defense requiring no parameter updates, external clean data, or auxiliary generation. TIGS leverages the observation that successful backdoor triggers consistently induce localized attention collapse within the semantic content region. Operating entirely within the native forward pass, TIGS first performs content-aware tail-risk screening to identify suspicious attention heads and rows using sample-internal signals. It then applies intrinsic geometric smoothing: a weak content-domain correction preserves semantic anchoring, while a stronger full-row contraction disrupts trigger-dominant routing. Finally, a controlled full-row write-back reconstructs the attention matrix to ensure inference stability. Extensive evaluations demonstrate that TIGS substantially suppresses attack success rates while strictly preserving clean reasoning and open-ended semantic consistency. Crucially, this favorable security-utility-latency equilibrium persists across diverse architectures, including dense, reasoning-oriented, and sparse mixture-of-experts models. By structurally disrupting adversarial routing with marginal latency overhead, TIGS establishes a highly practical, deployment-ready defense standard for state-of-the-art LLMs.

Large Language Models (LLMs), despite their impressive capabilities across domains, have been shown to be vulnerable to backdoor attacks. Prior backdoor strategies predominantly operate at the token level, where an injected trigger causes the model to generate a specific target word, choice, or class (depending on the task). Recent advances, however, exploit the long-form reasoning tendencies of modern LLMs to conduct reasoning-level backdoors: once triggered, the victim model inserts one or more malicious reasoning steps into its chain-of-thought (CoT). These attacks are substantially harder to detect, as the backdoored answer remains plausible and consistent with the poisoned reasoning trajectory. Yet, defenses tailored to this type of backdoor remain largely unexplored. To bridge this gap, we propose Critical-CoT, a novel defense mechanism that conducts a two-stage fine-tuning (FT) process on LLMs to develop critical thinking behaviors, enabling them to automatically identify potential backdoors and refuse to generate malicious reasoning steps. Extensive experiments across multiple LLMs and datasets demonstrate that Critical-CoT provides strong robustness against both in-context learning-based and FT-based backdoor attacks. Notably, Critical-CoT exhibits strong cross-domain and cross-task generalization. Our code is available at hthttps://github.com/tuanvu171/Critical-CoT.

We address jailbreaks, backdoors, and unlearning for large language models (LLMs). Unlike prior work, which trains LLMs based on their actions when given malign instructions, our method specifically trains the model to change how it interprets instructions. Our method, Latent Instruction Representation Alignment (LIRA), greatly improves generalization. We further boost generalization through an internally adversarial training algorithm. Our methods block over 99% of PEZ jailbreak attacks; remove a challenging insecure code backdoor; and achieve optimal forgetting on WMDP cyber with negligible loss of benign capabilities.

Multimodal large language models have become an important infrastructure for unified processing of visual and linguistic tasks. However, such models are highly susceptible to backdoor implantation during supervised fine-tuning and will steadily output the attacker's predefined harmful responses once a specific trigger pattern is activated. The core challenge of backdoor defense lies in suppressing attack success under low poisoning ratios while preserving the model's normal generation ability. These two objectives are inherently conflicting. Strong suppression often degrades benign performance, whereas weak regularization fails to mitigate backdoor behaviors. To this end, we propose a unified defense framework based on patch augmentation and cross-view regularity, which simultaneously constrains the model's anomalous behaviors in response to triggered patterns from both the feature representation and output distribution levels. Specifically, patch-level data augmentation is combined with cross-view output difference regularization to exploit the fact that backdoor responses are abnormally invariant to non-semantic perturbations and to proactively pull apart the output distributions of the original and perturbed views, thereby significantly suppressing the success rate of backdoor triggering. At the same time, we avoid over-suppression of the model during defense by imposing output entropy constraints, ensuring the quality of normal command generation. Experimental results across three models, two tasks, and six attacks show that our proposed defense method effectively reduces the attack success rate while maintaining a high level of normal text generation capability. Our work enables the secure, controlled deployment of large-scale multimodal models in realistic low-frequency poisoning and covert triggering scenarios.

The integration of large language models (LLMs) into robotic control pipelines enables natural language interfaces that translate user prompts into executable commands. However, this digital-to-physical interface introduces a critical and underexplored vulnerability: structured backdoor attacks embedded during fine-tuning. In this work, we experimentally investigate LoRA-based supply-chain backdoors in LLM-mediated ROS2 robotic control systems and evaluate their impact on physical robot execution. We construct two poisoned fine-tuning strategies targeting different stages of the command generation pipeline and reveal a key systems-level insight: back-doors embedded at the natural-language reasoning stage do not reliably propagate to executable control outputs, whereas backdoors aligned directly with structured JSON command formats successfully survive translation and trigger physical actions. In both simulation and real-world experiments, backdoored models achieve an average Attack Success Rate of 83% while maintaining over 93% Clean Performance Accuracy (CPA) and sub-second latency, demonstrating both reliability and stealth. We further implement an agentic verification defense using a secondary LLM for semantic consistency checking. Although this reduces the Attack Success Rate (ASR) to 20%, it increases end-to-end latency to 8-9 seconds, exposing a significant security-responsiveness trade-off in real-time robotic systems. These results highlight structural vulnerabilities in LLM-mediated robotic control architectures and underscore the need for robotics-aware defenses for embodied AI systems.

Backdoor attacks compromise model reliability by using triggers to manipulate outputs. Trigger inversion can accurately locate these triggers via a generator and is therefore critical for backdoor defense. However, the discrete nature of text prevents existing noise-based trigger generator from being applied to nature language processing (NLP). To overcome the limitations, we employ the rich knowledge embedded in large language models (LLMs) and propose a Backdoor defender powered by LLM Trigger Generator, termed BadLLM-TG. It is optimized through prompt-driven reinforcement learning, using the victim model's feedback loss as the reward signal. The generated triggers are then employed to mitigate the backdoor via adversarial training. Experiments show that our method reduces the attack success rate by 76.2% on average, outperforming the second-best defender by 13.7.

Backdoor attacks pose severe security threats to large language models (LLMs), where a model behaves normally under benign inputs but produces malicious outputs when a hidden trigger appears. Existing backdoor removal methods typically assume prior knowledge of triggers, access to a clean reference model, or rely on aggressive finetuning configurations, and are often limited to classification tasks. However, such assumptions fall apart in real-world instruction-tuned LLM settings. In this work, we propose a new framework for purifying instruction-tuned LLM without any prior trigger knowledge or clean references. Through systematic sanity checks, we find that backdoor associations are redundantly encoded across MLP layers, while attention modules primarily amplify trigger signals without establishing the behavior. Leveraging this insight, we shift the focus from isolating specific backdoor triggers to cutting off the trigger-behavior associations, and design an immunization-inspired elimination approach: by constructing multiple synthetic backdoored variants of the given suspicious model, each trained with different malicious trigger-behavior pairs, and contrasting them with their clean counterparts. The recurring modifications across variants reveal a shared "backdoor signature"-analogous to antigens in a virus. Guided by this signature, we neutralize highly suspicious components in LLM and apply lightweight finetuning to restore its fluency, producing purified models that withstand diverse backdoor attacks and threat models while preserving generative capability.

Recent intelligent systems integrate powerful Large Language Models (LLMs) through APIs, but their trustworthiness may be critically undermined by targeted attacks like backdoor and prompt injection attacks, which secretly force LLMs to generate specific malicious sequences. Existing defensive approaches for such threats typically rely on high access rights, impose prohibitive costs, and hinder normal inference, rendering them impractical for real-world scenarios. To solve these limitations, we introduce DualSentinel, a lightweight and unified defense framework that can accurately and promptly detect the activation of targeted attacks alongside the LLM generation process. We first identify a characteristic of compromised LLMs, termed Entropy Lull: when a targeted attack successfully hijacks the generation process, the LLM exhibits a distinct period of abnormally low and stable token probability entropy, indicating it is following a fixed path rather than making creative choices. DualSentinel leverages this pattern by developing an innovative dual-check approach. It first employs a magnitude and trend-aware monitoring method to proactively and sensitively flag an entropy lull pattern at runtime. Upon such flagging, it triggers a lightweight yet powerful secondary verification based on task-flipping. An attack is confirmed only if the entropy lull pattern persists across both the original and the flipped task, proving that the LLM's output is coercively controlled. Extensive evaluations show that DualSentinel is both highly effective (superior detection accuracy with near-zero false positives) and remarkably efficient (negligible additional cost), offering a truly practical path toward securing deployed LLMs. The source code can be accessed at https://doi.org/10.5281/zenodo.18479273.

LoRA adapters let users fine-tune large language models (LLMs) efficiently. However, LoRA adapters are shared through open repositories like Hugging Face Hub \citep{huggingface_hub_docs}, making them vulnerable to backdoor attacks. Current detection methods require running the model with test input data -- making them impractical for screening thousands of adapters where the trigger for backdoor behavior is unknown. We detect poisoned adapters by analyzing their weight matrices directly, without running the model -- making our method trigger-agnostic. For each attention projection (Q, K, V, O), our method extracts five spectral statistics from the low-rank update $ΔW$, yielding a 20-dimensional signature for each adapter. A logistic regression detector trained on this representation separates benign and poisoned adapters across three model families -- Llama-3.2-3B~\citep{llama3}, Qwen2.5-3B~\citep{qwen25}, and Gemma-2-2B~\citep{gemma2} -- on unseen test adapters drawn from instruction-following, reasoning, question-answering, code, and classification tasks. Across all three architectures, the detector achieves 100% accuracy.

As Large Language Models (LLMs) are increasingly integrated into software development workflows, they also become prime targets for adversarial attacks. Among these, backdoor attacks are a significant threat, allowing attackers to manipulate model outputs through hidden triggers embedded in training data. Detecting such backdoors remains a challenge, and one promising approach is the use of Spectral Signature defense methods that identify poisoned data by analyzing feature representations through eigenvectors. While some prior works have explored Spectral Signatures for backdoor detection in neural networks, recent studies suggest that these methods may not be optimally effective for code models. In this paper, we revisit the applicability of Spectral Signature-based defenses in the context of backdoor attacks on code models. We systematically evaluate their effectiveness under various attack scenarios and defense configurations, analyzing their strengths and limitations. We found that the widely used setting of Spectral Signature in code backdoor detection is often suboptimal. Hence, we explored the impact of different settings of the key factors. We discovered a new proxy metric that can more accurately estimate the actual performance of Spectral Signature without model retraining after the defense. Our findings contribute to a deeper understanding of the configurations of Spectral Signature defenses in securing code models and offer insights into potential improvements for future research.

Large Language Models (LLMs) are vulnerable to attacks like prompt injection, backdoor attacks, and adversarial attacks, which manipulate prompts or models to generate harmful outputs. In this paper, departing from traditional deep learning attack paradigms, we explore their intrinsic relationship and collectively term them Prompt Trigger Attacks (PTA). This raises a key question: Given a prompt, can we tell whether a hidden trigger is steering the model's behavior? We propose UniGuardian, to the best of our knowledge the first training-free LLM detector to jointly detect successfully activated prompt injection, backdoor, and adversarial attacks without knowing the attack type. Its shared mechanism measures how structured prompt perturbations shift the model's output distribution. Additionally, we introduce a single-forward strategy to optimize the detection pipeline, enabling simultaneous attack detection and text generation within a shared batched forward pass at each decoding step. Our experiments confirm that UniGuardian accurately and efficiently identifies trigger-activated prompts in LLMs.