Knowledge distillation (KD) is a vital technique for deploying deep neural networks (DNNs) on resource-constrained devices by transferring knowledge from large teacher models to lightweight student models. While teacher models from third-party platforms may undergo security verification (e.g., backdoor detection), we uncover a novel and critical threat: distillation-conditional backdoor attacks (DCBAs). DCBA injects dormant and undetectable backdoors into teacher models, which become activated in student models via the KD process, even with clean distillation datasets. While the direct extension of existing methods is ineffective for DCBA, we implement this attack by formulating it as a bilevel optimization problem and proposing a simple yet effective method (i.e., SCAR). Specifically, the inner optimization simulates the KD process by optimizing a surrogate student model, while the outer optimization leverages outputs from this surrogate to optimize the teacher model for implanting the conditional backdoor.

Backdoor attack is a common threat to deep neural networks.

To mitigate the threat, existing solutions attempted to search for backdoor triggers, which can be time-consuming when handling a large search space.

As Large Language Models (LLMs) gain traction across critical domains, ensuring secure and trustworthy training processes has become a major concern. Backdoor attacks, where malicious actors inject hidden triggers into training data, are particularly insidious and difficult to detect. Existing post-training verification solutions like Proof-of-Learning are impractical for LLMs due to their requirement for full retraining, lack of robustness against stealthy manipulations, and inability to provide early detection during training. Early detection would significantly reduce computational costs. To address these limitations, we introduce Proof-of-Training Steps, a verification protocol that enables an independent auditor (Alice) to confirm that an LLM developer (Bob) has followed the declared training recipe, including data batches, architecture, and hyperparameters. By analyzing the sensitivity of the LLMs' language modeling head (LM-Head) to input perturbations, our method can expose subtle backdoor injections or deviations in training. Even with backdoor triggers in up to 10 percent of the training data, our protocol significantly reduces the attacker's ability to achieve a high attack success rate (ASR). Our method enables early detection of attacks at the injection step, with verification steps being 3x faster than training steps. Our results highlight the protocol's potential to enhance the accountability and security of LLM development, especially against insider threats.

To mitigate the threat, existing solutions attempted to search for backdoor triggers, which can be time-consuming when handling a large search space.

We draw inspiration from microsaccades, tiny involuntary eye movements that reveal hidden dynamics of human perception, to propose an analogous probing method for large language models (LLMs). Just as microsaccades expose subtle but informative shifts in vision, we show that lightweight position encoding perturbations elicit latent signals that indicate model misbehaviour. Our method requires no fine-tuning or task-specific supervision, yet detects failures across diverse settings including factuality, safety, toxicity, and backdoor attacks. Experiments on multiple state-of-the-art LLMs demonstrate that these perturbation-based probes surface misbehaviours while remaining computationally efficient. These findings suggest that pretrained LLMs already encode the internal evidence needed to flag their own failures, and that microsaccade-inspired interventions provide a pathway for detecting and mitigating undesirable behaviours.

Knowledge distillation (KD) is a vital technique for deploying deep neural networks (DNNs) on resource-constrained devices by transferring knowledge from large teacher models to lightweight student models. While teacher models from third-party platforms may undergo security verification (\eg, backdoor detection), we uncover a novel and critical threat: distillation-conditional backdoor attacks (DCBAs). DCBA injects dormant and undetectable backdoors into teacher models, which become activated in student models via the KD process, even with clean distillation datasets. While the direct extension of existing methods is ineffective for DCBA, we implement this attack by formulating it as a bilevel optimization problem and proposing a simple yet effective method (\ie, SCAR). Specifically, the inner optimization simulates the KD process by optimizing a surrogate student model, while the outer optimization leverages outputs from this surrogate to optimize the teacher model for implanting the conditional backdoor. Our SCAR addresses this complex optimization utilizing an implicit differentiation algorithm with a pre-optimized trigger injection function. Extensive experiments across diverse datasets, model architectures, and KD techniques validate the effectiveness of our SCAR and its resistance against existing backdoor detection, highlighting a significant yet previously overlooked vulnerability in the KD process. Our code is available at https://github.com/WhitolfChen/SCAR.

Institutions with limited data and computing resources often outsource model training to third-party providers in a semi-honest setting, assuming adherence to prescribed training protocols with pre-defined learning paradigm (e.g., supervised or semi-supervised learning). However, this practice can introduce severe security risks, as adversaries may poison the training data to embed backdoors into the resulting model. Existing detection approaches predominantly rely on statistical analyses, which often fail to maintain universally accurate detection accuracy across different learning paradigms. To address this challenge, we propose a unified backdoor detection framework in the semi-honest setting that exploits cross-examination of model inconsistencies between two independent service providers. Specifically, we integrate central kernel alignment to enable robust feature similarity measurements across different model architectures and learning paradigms, thereby facilitating precise recovery and identification of backdoor triggers. We further introduce backdoor fine-tuning sensitivity analysis to distinguish backdoor triggers from adversarial perturbations, substantially reducing false positives. Extensive experiments demonstrate that our method achieves superior detection performance, improving accuracy by 5.4%, 1.6%, and 11.9% over SoTA baselines across supervised, semi-supervised, and autoregressive learning tasks, respectively. Notably, it is the first to effectively detect backdoors in multimodal large language models, further highlighting its broad applicability and advancing secure deep learning.

Visual prompting (VP) is a new technique that adapts well-trained frozen models for source domain tasks to target domain tasks. This study examines VP's benefits for black-box model-level backdoor detection. The visual prompt in VP maps class subspaces between source and target domains. We identify a misalignment, termed class subspace inconsistency, between clean and poisoned datasets. Deep neural networks (DNNs) are commonly used in complex applications but require extensive computational power, leading to significant costs. However, DNNs can include backdoors (Gu et al., 2017; Liu et al., 2018b; Tang et al., 2021; Qi et al., 2023b; Nguyen & Tran, 2021; Chen et al., 2017), which manipulate model responses to inputs with specific triggers (like certain pixel patterns) while functioning correctly on other inputs. In backdoor attacks, attackers embed these triggers in the training data, leading the model to associate the trigger with a particular outcome and misclassify inputs containing it. Black-box backdoor detection, which uses only blackbox queries to the suspicious model (i.e., the model to be inspected), is gaining attention.